Content closed-loop interaction system based on large language model

Abstract

The application discloses a content closed-loop interaction system based on a large language model and belongs to the technical field of large language model interaction systems. The system comprises a field feature construction unit, a double-channel feature extraction unit, an interaction content generation module and a correction module. The field feature construction unit comprises a data acquisition calibration unit for acquiring industry data. The data acquisition calibration unit is connected with an industry terminology library for storing industry standards and a three-dimensional association library of associated terms. Each associated term comprises a field feature word, elements of a certain subset and functions. The double-channel feature extraction unit is in communication connection with an input processing module. The double-channel feature extraction unit comprises a general semantic channel and a field feature channel composed of a large language model. The content closed-loop interaction system based on the large language model can more accurately extract and express interaction content and adopts closed-loop data management to realize efficient, accurate and reliable intelligent interaction.

Description

Technical Field

[0001] This invention specifically relates to a content closed-loop interactive system based on a large language model, belonging to the technical field of large language model interactive systems. Background Technology

[0002] With the rapid development of large language model technology, its application in fields such as natural language processing and intelligent dialogue is becoming increasingly widespread. For example, Chinese Patent Publication No. CN119378572B discloses a digital consecutive interpreting method, device, and equipment based on a large language model. This method utilizes pre-stored benchmark semantic interaction vectors and digital annotation indication strategies for digital consecutive interpreting, improving accuracy and avoiding errors in digital conversion and semantic understanding that may occur with manual interpretation. The entire process is highly automated, improving interpreting efficiency and meeting the needs of rapid business communication. Another example is Chinese Patent Publication No. CN120108396A, which discloses a real-time digital human interaction system and its AI voice dialogue chip circuit system. The system's large model module is based on a large LLM model finely tuned with a large amount of data, API interfaces, or workflow agents to achieve functions such as speech recognition, text-to-speech, lip-sync, real-time audio and video interaction, and real-time translation. However, existing large language model systems have the following problems:

[0003] 1. Lack of an effective feedback mechanism makes it impossible to continuously optimize the output content based on application interaction. The interaction process is mostly one-way output, making it difficult to form an effective learning loop.

[0004] 2. Unable to understand the professional terminology of different industries and output professional interactive data;

[0005] 3. When large language models are used in various real-world scenarios involving reasoning problems, errors in the reasoning process can lead to unreliable results, resulting in unreliable text. In existing technologies, the credibility of text can only be verified based on one's own experience, making it difficult to verify whether the generated content is true and reliable. Summary of the Invention

[0006] To address the aforementioned issues, this invention proposes a content-closed-loop interactive system based on a large language model. This system enables more precise extraction and representation of interactive content and employs closed-loop data management to achieve efficient, accurate, and reliable intelligent interaction.

[0007] The content closed-loop interactive system based on a large language model of the present invention includes:

[0008] The domain feature construction unit includes a data acquisition and calibration unit for collecting industry data. The data acquisition and calibration unit is connected to a three-dimensional association library for storing industry standard terminology and related terms. Each related term includes a domain feature word, elements of a subset, and functions.

[0009] A dual-channel feature extraction unit is communicatively connected to the input processing module. This dual-channel feature extraction unit constitutes a dual-channel attention mechanism, including a general semantic channel and a domain feature channel constructed from a large language model. The general semantic channel receives the input text, passes it through the Transformer layer of the large language model's LLM, and extracts general language features. The domain feature channel communicates with the content construction unit to query whether the input text contains domain feature words and their similarity to those words, thereby obtaining related or similar related term content.

[0010] An interactive content generation module analyzes and determines interactive information corresponding to the input text based on general language features and domain features; the interactive information is content closed-loop data composed of result data and analysis data; the analysis data is provided by an industry terminology database and a three-dimensional association database.

[0011] The error correction module communicates with the interactive content generation module. After obtaining the result data and analysis data, and determining that an information deviation has occurred, the error correction module manually revises the result data or analysis data to obtain revised data. It then uploads the general language features and domain-specific features as search terms and the revised data as bound revised data to the revision database. When only the result data or only the analysis data is modified, the unmodified data remains in its initial state, and the modified data is individually labeled. When the input text overlaps with the search terms in the revision database, the revised data content is output first. When the result data or analysis data is modified, the modified data is synchronously updated and supplemented into the three-dimensional association database.

[0012] The content closed-loop interactive system based on a large language model of the present invention takes the medical and health field as an example, and describes in detail the application of the domain feature construction unit, dual-channel feature extraction unit, interactive content generation module and correction module in real-world scenarios; it demonstrates how to achieve accurate interactive content generation and correction through this technical architecture.

[0013] Data acquisition and calibration unit:

[0014] We collect industry data in the healthcare field, including medical records, medical literature, and clinical guidelines. After data collection, we use both manual annotation and automated tools to calibrate the data, ensuring accuracy and consistency. Our system includes an industry terminology database and a three-dimensional association database: The industry terminology database stores standardized medical terms such as "hypertension" and "diabetes," along with authoritative definitions and classifications. The three-dimensional association database includes each related term as follows: domain-specific keywords (e.g., "hypertension"); subset elements (e.g., "systolic blood pressure," "diastolic blood pressure," "complications"); and functions (e.g., "diagnostic criteria," "treatment plans," "preventive measures").

[0015] Dual-channel feature extraction unit:

[0016] Input processing module: Receives text input from the user, such as: "What are the treatment options for hypertension?"

[0017] General Semantic Channel: Extract general language features through the Transformer layer of a large language model to understand the semantics and intent of user questions;

[0018] Domain Feature Channel: Query whether the input text contains domain feature words (such as "hypertension") and calculate its similarity with the domain feature words; retrieve related terms related to "hypertension" from the three-dimensional association library, such as: "treatment plan" including "drug treatment" and "lifestyle intervention".

[0019] Interactive content generation module:

[0020] Common language feature: Identifies user questions as "treatment plan" queries;

[0021] Domain characteristics: Retrieve a subset of "treatment plans" for "hypertension" from a 3D relational database;

[0022] The specific details of generating interactive information are as follows:

[0023] Results data: Directly answering user questions, such as: "Treatment options for hypertension include medication (such as ACE inhibitors and calcium channel blockers) and lifestyle interventions (such as a low-sodium diet and regular exercise)."

[0024] Data analysis: Supplement the industry terminology database and the three-dimensional association database with detailed analysis, such as: "According to the 'Guidelines for the Prevention and Treatment of Hypertension in China,' drug treatment should be selected according to the individual patient's condition, and lifestyle intervention is the basic treatment method."

[0025] The working process of the correction module is as follows:

[0026] Information bias detection: If the "treatment plan" output by the system is missing the "traditional Chinese medicine treatment" option, the user can manually revise it through the bias correction module;

[0027] Revised data:

[0028] Users added "Traditional Chinese Medicine Treatment" to the "Treatment Plan" and marked it as supplementary content;

[0029] The revised data (“Traditional Chinese Medicine Treatment”) is linked with common language features (“Treatment Plan”) and domain feature terms (“Hypertension”) and uploaded to the revision database;

[0030] Data Update:

[0031] The revised data is synchronously updated to the "hypertension" related terms in the three-dimensional association database;

[0032] When users query the same question in the future, the system will prioritize outputting the revised content.

[0033] Closed-loop data output:

[0034] The system ultimately outputs closed-loop data, including:

[0035] Results data: "Treatment options for hypertension include medication, lifestyle interventions, and traditional Chinese medicine."

[0036] Data analysis: "Traditional Chinese medicine treatment is suitable for some patients and should be carried out under the guidance of a doctor."

[0037] The linkage mechanism of related databases ensures the priority and consistency of query results.

[0038] Furthermore, the interactive content generation module performs the following analysis process based on general language features and domain features: The large language model processes the input text to obtain feature data of general language features and domain features. Based on the feature data, it determines explicit and implicit intentions. Explicit intentions are those that can be directly obtained through feature data, while implicit intentions are emotional intentions or intentions that cannot be directly derived from explicit intentions and require reasoning. The reasoning process is as follows: Each feature element in the general language features and domain features is used as a reasoning logic neuron, where the feature elements of the domain features are domain feature words and their own functions. The large language model uses the reasoning logic neurons as basic elements, combines each basic element, and performs semantic reasoning to obtain N sets of reasoning combinations. The N sets of reasoning combinations are then summarized, and those with intent similarity reaching a set value are included in the same combination. The number of combinations is counted and arranged from high to low. Multiple sets of reasoning results are displayed through a system-initiated query method, guiding the selection of a certain reasoning result, and the result is used as the implicit intention. Finally, the large language model outputs the result data that matches the explicit or implicit intention, as well as the analysis data corresponding to the structural data.

[0039] Furthermore, the result data consists of a set of one or more results, and the analysis data provides inference evidence for each subset in the set. The inference evidence is the intent, the inference relationship edge between the intent and a subset or an element of the subset in the three-dimensional association library, and the inference relationship edge between the intent and an industry standard in the industry terminology library.

[0040] Furthermore, when the interactive content generation module generates interactive information, it obtains general language features and domain features through a dual-channel feature extraction unit and calculates their hit probability with the search term. When the overlap between the searched domain features and the domain features of the search term reaches a set value, it is considered a hit; or a weight is set for each domain feature word of the search term, and when the weight value reaches the set value, it is considered a hit; or a weight is set for general language features and domain features, first searching for the overlap between the input text and the general language features and domain features of the search term, and then calculating the total weight value. When the weight value reaches the set value, it is considered a hit; after the input text and the search term are considered a hit, the interactive content generation module outputs the revised data content corresponding to the search term.

[0041] Furthermore, the weight of domain feature words in the search terms is set as follows: After the correction module completes the revision data revision, it enters the revision data binding process. At this time, the weight of domain feature words in the search terms can be set.

[0042] Furthermore, when generating interactive information, the interactive content generation module balances general semantics and domain feature data through a gating mechanism.

[0043] Furthermore, after the correction module modifies the analyzed data, the modified data is used to optimize the large language model through incremental updates of a low-rank matrix.

[0044] Furthermore, the dual-channel feature extraction unit is connected to a training module, which acquires a domain training set; the dual-channel feature extraction unit acquires the domain training set, performs general language features and domain features extraction on the domain training set, and the large language model acquires the intent and reasoning relationship edge corresponding to each piece of data in the domain training set.

[0045] Furthermore, when the dual-channel feature extraction unit acquires domain feature words, it simultaneously generates similar word entries, enriching the number of domain feature words acquired by the dual-channel feature extraction unit and improving the comprehensiveness of subsequent reasoning combinations.

[0046] Compared with existing technologies, the content closed-loop interactive system based on a large language model of the present invention adopts a dual-channel attention mechanism to separate general semantics and domain features, realize dual-channel feature extraction, and combine general semantics and domain features to ensure the professionalism and accuracy of the answers. Furthermore, it balances the dynamic fusion of industry knowledge through a gating mechanism, enabling more precise extraction and expression of interactive content. During the interaction process, it provides an analysis process for the results, increasing the credibility of the large language model content interaction. The system is also scalable, dynamically updated through an industry terminology database and a three-dimensional association database, supporting multi-domain expansion or transforming a general interactive system into a precise, industry-specific interactive system. User feedback allows for direct correction of system output, forming a continuously optimized closed loop. In addition, the linkage mechanism between the revision database, the three-dimensional revision database, and the three-dimensional association database ensures the priority and consistency of query results. Ultimately, it achieves efficient, accurate, and reliable intelligent interaction. Attached Figure Description

[0047] Figure 1 This is a block diagram of the overall structure of the content closed-loop interactive system of the present invention.

[0048] Figure 2 This is a schematic diagram illustrating the specific interactions of each module in the content closed-loop interactive system of the present invention.

[0049] Figure 3 This is a schematic diagram of the analysis process of the interactive content generation module of the present invention based on common language features and domain features.

[0050] Figure 4 This is a schematic diagram of the process of generating interactive information by the interactive content generation module of the present invention. Detailed Implementation

[0051] like Figures 1 to 4 The content-closed-loop interactive system based on a large language model, as shown, includes:

[0052] The domain feature construction unit includes a data acquisition and calibration unit for collecting industry data. The data acquisition and calibration unit is connected to a three-dimensional association library for storing industry standard terminology and related terms. Each related term includes a domain feature word, elements of a subset, and functions.

[0053] A dual-channel feature extraction unit is communicatively connected to the input processing module. This dual-channel feature extraction unit constitutes a dual-channel attention mechanism, including a general semantic channel and a domain feature channel constructed from a large language model. The general semantic channel receives the input text, passes it through the Transformer layer of the large language model's LLM, and extracts general language features. The domain feature channel communicates with the content construction unit to query whether the input text contains domain feature words and their similarity to those words, thereby obtaining related or similar related term content.

[0054] An interactive content generation module analyzes and determines interactive information corresponding to the input text based on general language features and domain features; the interactive information is content closed-loop data composed of result data and analysis data; the analysis data is provided by an industry terminology database and a three-dimensional association database.

[0055] The error correction module communicates with the interactive content generation module. After obtaining the result data and analysis data, and determining that an information deviation has occurred, the error correction module manually revises the result data or analysis data to obtain revised data. It then uploads the general language features and domain-specific features as search terms and the revised data as bound revised data to the revision database. When only the result data or only the analysis data is modified, the unmodified data remains in its initial state, and the modified data is individually labeled. When the input text overlaps with the search terms in the revision database, the revised data content is output first. When the result data or analysis data is modified, the modified data is synchronously updated and supplemented into the three-dimensional association database.

[0056] The interactive content generation module performs the following analysis based on general language features and domain features: The large language model processes the input text to obtain feature data of general language features and domain features. Based on the feature data, it determines explicit and implicit intentions. Explicit intentions are those that can be directly obtained from the feature data, while implicit intentions are emotional intentions or intentions that cannot be directly derived from explicit intentions and require inference. The inference process is as follows: Each feature element in the general language features and domain features is used as an inference logic neuron, where the feature elements of the domain features are domain feature words and their own functions. The large language model uses the inference logic neurons as basic elements, combines each basic element, and performs semantic inference to obtain N sets of inference combinations. The N sets of inference combinations are then summarized, and those with intent similarity reaching a set value are included in the same combination. The number of combinations is counted and arranged from high to low. Multiple inference results are displayed through a system-initiated query method, guiding the selection of a certain inference result, and the result is used as the implicit intention. Finally, the large language model outputs the matching result data and the analysis data corresponding to the structural data based on the explicit or implicit intention.

[0057] For example, a content-based closed-loop interaction system based on a large language model can be used for intent recognition and response generation in an intelligent customer service system:

[0058] Scenario Description: Suppose a user enters the text: "My order shows as shipped, but the logistics information hasn't been updated. Can you help me check it?"

[0059] The system needs to identify explicit intent (querying logistics status) and implicit intent (which may imply anxiety about logistics delays or a need to expedite the process).

[0060] The feature extraction stage is as follows:

[0061] Common language features: Keywords: order, shipped, logistics information, query;

[0062] Sentence structure: interrogative sentence + problem description;

[0063] Emotional tendency: Neutral to anxious ("No updates for a long time" implies dissatisfaction);

[0064] Domain characteristics (domain terms): orders, logistics, shipping;

[0065] Related functions: Logistics query API and order status interface;

[0066] The intention reasoning phase is as follows:

[0067] Explicit intent: Directly determine user needs through the keyword "query logistics information";

[0068] Implicit Intent Reasoning: The model generates the following hypotheses through the combination of reasoning logic neurons:

[0069] Combination 1: Users want to proactively follow up on the reasons for logistics delays (implying an urging);

[0070] Basis: The sentiment phrase "has not been updated" + the domain function "logistics anomaly handling";

[0071] Option 2: Users only need to confirm the logistics status (no implicit requirements);

[0072] Basis: Explicit intent directly overrides;

[0073] The system statistically found that combination 1 had a high degree of intent similarity (due to strong emotional characteristics), and included it in the implicit intent candidate;

[0074] Interactive verification: The system proactively asks the user to clarify the implicit intent through interactive means.

[0075] "We've detected that you may have questions about the shipping delay. Would you like our assistance in contacting the logistics company?"

[0076] After the user selects "yes", the implicit intent is confirmed as "urging logistics to follow up".

[0077] Output phase: Explicit intent response, calling the logistics API to return the latest status: "Current logistics information has been updated, the package is expected to arrive tomorrow."

[0078] Implicit Intent Response: Supplementing with reassuring analysis data: "Due to the recent peak in logistics causing delays, we will prioritize your case and will notify you of the progress via SMS."

[0079] For example, a content-based closed-loop interactive system based on a large language model can be used in an equipment maintenance decision support system for implicit intent analysis.

[0080] Scenario description: A centrifugal pump in a chemical plant is experiencing "periodic fluctuations in outlet pressure + abnormal current". The maintenance personnel described it as: "The pressure gauge pointer of the P-203A pump is swinging regularly, and the motor current is fluctuating between 30-40A".

[0081] Construction of reasoning logic neurons:

[0082] General features:

[0083] Periodic fluctuations (feature elements) -- time pattern recognition (functional attributes);

[0084] Numerical jumps (feature elements) -- anomaly threshold detection (functional attributes);

[0085] Domain characteristics:

[0086] Centrifugal pump (P-203A) (characteristic element) -- fluid transport equipment (functional attribute);

[0087] Export pressure fluctuation (characteristic element) -- System resistance change indicator (functional attribute);

[0088] Motor current anomaly (feature element) -- load status monitoring (functional attribute);

[0089] Multi-dimensional reasoning combination: periodic fluctuations, numerical jumps, centrifugal pumps, outlet pressure fluctuations, abnormal motor current;

[0090] Generate reasoning combinations (actually semantic space computation); details are as follows:

[0091] The fluctuations in centrifugal pump outlet pressure are attributed to cavitation; abnormal motor current and periodic fluctuations are attributed to power supply interference; fluctuations in centrifugal pump outlet pressure and abnormal motor current are attributed to impeller wear; numerical jumps and abnormal motor current are attributed to inverter malfunctions; the similarity clustering results are shown in Table 1.

[0092]

[0093] Implicit Intent Recognition Interaction Flow:

[0094] The system proactively asks: Multiple possible causes have been detected. Please confirm:

[0095] 1. Has the medium been changed recently? (Related to cavitation);

[0096] 2. Has the last major overhaul occurred more than 8,000 hours ago? (Mechanical wear and tear);

[0097] 3. Have you observed the alarm lights on the control cabinet? (Electrical fault).

[0098] User selection: Maintenance personnel feedback: "We switched to high-viscosity raw materials last week". The system will: mark "We switched to high-viscosity raw materials last week" as an implicit intent; automatically eliminate electrical fault combinations.

[0099] Results data: Output handling plan: 1. Immediately reduce the pump speed to 80% of the rated value; 2. Check the pressure difference of the inlet filter.

[0100] Data analysis: 1. Industry standard SH / T 3538-2017: High viscosity media require a reduction in NPSHr value.

[0101] 2. Three-dimensional correlation library records: The feature word "viscosity mutation" is associated with "cavitation risk coefficient increased by 40%"; Historical case CT-2023-045: Similar phenomena were solved by adding an inducer wheel.

[0102] For example, some implicit intentions and triggering features are shown in Table 2:

[0103]

[0104] The result data consists of one or more results forming a set. The analysis data provides inference evidence for each subset in the set. The inference evidence is the intent, the inference relationship edge between the intent and a subset or an element of the subset in the three-dimensional association library, and the inference relationship edge between the intent and an industry standard in the industry terminology library.

[0105] Taking vibration spectrum analysis in bearing fault diagnosis as an example:

[0106] The text entered is: "The vibration spectrum of the centrifugal pump bearing shows that high-frequency harmonics account for 18%, accompanied by an increase in temperature. What could be the possible cause?"

[0107] 3D related library query: Domain feature words: abnormal bearing vibration;

[0108] Subset elements: high-frequency harmonics (vibration spectrum characteristics) and temperature rise (auxiliary characteristics);

[0109] Cause of failure: bearing wear, lubrication failure

[0110] Functions: Trigger warning thresholds (e.g., high-frequency harmonics > 15%) and associate with maintenance work orders;

[0111] The process of constructing the reasoning relation is as follows:

[0112] meaning Figure 1 High-frequency harmonics account for 18% → bearing wear

[0113] Reasoning relation edge: Associated with the 3D association library: High frequency harmonics → Bearing wear (supporting edge, weight 0.9) Associated with the industry terminology library: ISO13374-2 vibration standard (referencing the standard "High frequency harmonics are a typical characteristic of bearing wear").

[0114] meaning Figure 2 Increased temperature leads to lubrication failure.

[0115] Reasoning relation edge: Associated with the 3D association library: Temperature rise → Lubrication failure (support edge, weight 0.7) Associated with the industry terminology library: FMEA failure mode (referring to "insufficient lubrication leads to frictional temperature rise").

[0116] Interactive content generation:

[0117] Results data: "Possible causes: 1. Bearing wear (probability 85%); 2. Lubrication failure (probability 60%). It is recommended to check the condition of the bearing grease first."

[0118] Data analysis: The proportion of high-frequency harmonics is 18%, which conforms to the bearing wear characteristics of ISO13374-2 (referencing standard clause 4.2.1).

[0119] Temperature rise is associated with lubrication failure modes in FMEA (see case number ZWEM-2021-016).

[0120] Correction and optimization: Based on feedback from engineers, actual disassembly and inspection revealed that the bearing balls had peeled off, which was not simply due to wear.

[0121] Revision operation: Added a new 3D related library entry: ball bearing peeling (features: high frequency harmonics + intermittent impact waveform).

[0122] Update the inference relation edge: high frequency harmonics → ball spalling (weight 1.0), and bind it to the industry terminology library ISO10816-3 impact vibration standard.

[0123] Closed-loop effect: Subsequent similar inputs are automatically associated with the "ball peeling" feature, improving diagnostic accuracy.

[0124] When generating interactive information, the interactive content generation module obtains general language features and domain features through a dual-channel feature extraction unit and calculates their hit probability with the search term. When the overlap between the retrieved domain features and the domain features of the search term reaches a set value, it is considered a hit; or a weight is set for each domain feature word of the search term, and when the weight value reaches the set value, it is considered a hit; or a weight is set for general language features and domain features, first retrieving the overlap between the input text and the general language features and domain features of the search term, and then calculating the total weight value. When the weight value reaches the set value, it is considered a hit; after the input text and the search term are considered a hit, the interactive content generation module outputs the revised data content corresponding to the search term.

[0125] When the interactive content generation module generates interactive information, it calculates the probability of its match with the search term. This is applied as follows: for example, in the hit calculation of an intelligent fault diagnosis system:

[0126] Scenario Description: In the intelligent maintenance system for wind power equipment, the user inputs the text "abnormal vibration frequency of gearbox, accompanied by periodic clicking sound". The system needs to match the revised data in the historical maintenance case library and output the cause of the fault and the solution.

[0127] Data preparation:

[0128] Domain-specific keywords in the domain-specific keyword library: gearbox, vibration frequency, and periodic abnormal noise;

[0129] Related terms:

[0130] Entry 1: Gearbox → Subset Element: Bearing Wear → Function: Recommend replacing the bearing;

[0131] Entry 2: Vibration frequency → Subset element: Gear meshing misalignment → Function: Adjust gear clearance;

[0132] Revision Database: Search terms: General feature vibration anomaly + Domain feature gearbox, bound revision data:

[0133] Results: Bearing wear probability 85%; Analysis data: Needs confirmation in conjunction with temperature sensor data.

[0134] The hit calculation process is as follows:

[0135] Dual-channel feature extraction:

[0136] General Semantic Channel (LLM) for feature extraction: mechanical faults, vibration analysis;

[0137] Domain Feature Channel: Hit terms: Gearbox (weight 0.5), Vibration frequency (weight 0.3);

[0138] Similar terms: periodic abnormal noise → matching bearing wear (cosine similarity 0.7);

[0139] Weighting calculation: Input text features: General feature weight: 0.4; Domain feature weight: 0.6 (gearbox × 0.5 + vibration frequency × 0.3 + periodic abnormal noise × 0.7 × 0.2);

[0140] Search term features: General feature vibration anomaly weight 0.4 (match degree 0.8); Domain feature gearbox weight 0.6 (perfect match); Total weight = (0.4×0.8)+(0.6×1.0)=0.92>set threshold 0.85, hit completed; Output result: Fault cause: bearing wear (85% confidence level), recommended measures: replace bearing and check temperature sensor data.

[0141] The weighting of domain feature words for the search terms is set as follows: After the correction module completes the revision data revision, it enters the revision data binding process. At this time, the weighting of domain feature words for the search terms can be set.

[0142] When generating interactive information, the interactive content generation module uses a gating mechanism to balance general semantics and domain feature data. The gating mechanism's balancing calculation is as follows:

[0143] (1);

[0144] The gating signal is generated as follows:

[0145] (2);

[0146] in, For the gating mechanism balance calculation results, These are general semantic features (semantic features extracted by general semantic channel LLM). Domain features (domain features matched by the data acquisition and calibration unit); For the learnable parameter matrix, calculate the balanced weights of general semantic features and domain features; For feature splicing operations; for Activation functions; they enable dynamic weight allocation, and the closer the output value is to 1, the stronger the corresponding feature. The greater the contribution, the closer it is to 0, the more it inhibits. .

[0147] After the correction module modifies the analyzed data, the modified data is used to optimize the large language model through incremental updates of a low-rank matrix.

[0148] The dual-channel feature extraction unit is connected to a training module, which acquires a domain training set. The dual-channel feature extraction unit acquires the domain training set and extracts general language features and domain features from the domain training set. The large language model acquires the intent and reasoning relationship edges corresponding to each piece of data in the domain training set.

[0149] When the dual-channel feature extraction unit acquires domain feature words, it simultaneously generates similar entries, such as matching the related entry "abnormal bearing vibration", retrieving historical fault cases (such as "a certain historical record of a similar pump causing excessive vibration due to bearing wear"), and supplementing the similar entry "impeller imbalance".

[0150] Example 1:

[0151] A content-closed-loop interactive system based on a large language model is used for fault diagnosis of intelligent devices. The specific architecture includes:

[0152] Domain Feature Construction Unit: The domain feature construction unit includes a data acquisition and calibration unit, which interfaces with the Industrial Internet of Things (IIoT) platform to collect equipment sensor data (such as vibration, temperature, current, etc.) and label fault types (such as bearing wear, motor overload); an industry terminology library stores industry standard terms (such as ISO 13374 equipment condition monitoring standard, FMEA fault code classification); and an example of related terms in the three-dimensional association library:

[0153] Domain-specific keywords: "abnormal bearing vibration", subset elements: vibration spectrum characteristics (high-frequency harmonics), fault causes (insufficient lubrication, ball shedding), functions: triggering early warning thresholds, associating with maintenance work orders;

[0154] Domain-specific keywords: motor; subset: speed reducer, electric slide; element: power drive unit; function: to provide rotational power for equipment.

[0155] Dual-channel feature extraction unit:

[0156] Input text: Operator asked: "The vibration value of centrifugal pump #1 has suddenly increased. What could be the problem?"

[0157] General semantic channel: LLM extracts keywords "centrifugal pump" and "increased vibration" and identifies them as equipment anomaly diagnosis needs;

[0158] Domain Feature Channel: Match related terms for "abnormal bearing vibration", retrieve historical fault cases (such as "a historical record of a similar pump causing excessive vibration due to bearing wear"), and supplement similar terms such as "impeller imbalance";

[0159] Interactive content generation module:

[0160] Results data: "Possible causes: 1. Bearing wear or impeller imbalance, it is recommended to check the bearing condition first";

[0161] Data analysis: Referencing the "Vibration Spectrum Feature Comparison Table" in the 3D correlation library, mark the fault characteristics (such as "High-frequency harmonics proportion > 15%").

[0162] The correction module performs corrections, and the specific process is as follows:

[0163] Deviation scenario: The engineer reported that "actual disassembly and inspection revealed that the impeller blades were broken, not a bearing problem."

[0164] Revision operation: Manually modify the result data, add related terms for "impeller fracture" (such as fracture characteristics: low frequency vibration + metal debris), and bind the search terms "centrifugal pump, vibration, impeller".

[0165] Library Update: Revised data is synchronized to the "Centrifugal Pump Failure Modes" subset of the 3D related library, and the revised content will be output first for similar issues in the future.

[0166] Closed-loop optimization: When a new input "2# pump vibration accompanied by abnormal noise" is entered, the system automatically associates the "impeller fracture" feature in the revision library and adds the association rule "abnormal noise frequency is synchronized with impeller speed", forming dynamic knowledge iteration.

[0167] Example 2:

[0168] A content-closed-loop interactive system based on a large language model is used for intelligent medical question answering. The specific architecture includes:

[0169] The domain feature construction unit is constructed as follows:

[0170] Data Acquisition and Calibration Unit: By connecting to the hospital's HIS system, electronic medical record database, and authoritative medical journals (such as The Lancet), it automatically collects structured medical record data (such as ICD-10 disease codes) and unstructured medical literature.

[0171] Industry Terminology Database: Stores standardized medical terms (such as "myocardial infarction" corresponding to ICD-10 code I21), including WHO disease classifications, generic drug names, etc.

[0172] 3D Association Library: Examples of associated terms are as follows:

[0173] Domain-specific keywords: "Aspirin", subset elements: chemical formula C9H8O4, pharmacological mechanism (COX-1 inhibition), function: antiplatelet aggregation, indication (secondary prevention of coronary heart disease);

[0174] Dual-channel feature extraction unit:

[0175] Input text: Patient's question: "Will long-term use of aspirin damage my stomach?"

[0176] General semantic channel: Extract general features (such as "long-term" and "damages the stomach") through the Transformer layer of LLM and identify them as drug side effect consultation;

[0177] Domain Feature Channel: Match related terms for "aspirin", retrieve "risk of gastrointestinal mucosal damage" in pharmacological mechanisms, and supplement similar terms such as "side effects of NSAIDs".

[0178] Interactive content generation module:

[0179] Results data: "Long-term use of aspirin may increase the risk of gastric ulcers, and it is recommended to use a proton pump inhibitor (such as omeprazole) in combination."

[0180] Data analysis: Referencing the "drug interactions" subset from the 3D association database, the evidence level is marked (e.g., "Level A clinical guideline recommendation").

[0181] Correction module:

[0182] Deviation scenario: If users report that "the actual guideline recommends a dose of 75-100 mg / day", but the initial results did not mention a dose range;

[0183] Revision operation: Manually modify the analysis data, add the dosage clause of the "Chinese Expert Consensus on Antiplatelet Therapy", and bind the search term "aspirin + dosage";

[0184] Library Update: The revised data is synchronized to the "Medication Guidelines" subset of the 3D related library, and the revised content will be output first for subsequent similar questions;

[0185] Closed-loop optimization: When a new input is made, “Are enteric-coated aspirin tablets safer?”, the system automatically associates the dosage data in the revision library and adds the related entry “Enteric-coated tablets can reduce gastric irritation”.

[0186] The above embodiments are merely preferred embodiments of the present invention. Therefore, all equivalent changes or modifications made to the structure, features and principles described in the claims of the present invention are included within the scope of the present invention.

Claims

1. A content closed-loop interaction system based on a large language model, characterized in that: Comprise: Domain feature construction unit, the domain feature construction unit includes a data collection calibration unit for collecting industry data, the data collection calibration unit is connected for storing industry standard industry terminology library and associated three-dimensional associated library, each associated word includes domain feature word, some subset of elements and functions; Dual-channel feature extraction unit, the dual-channel feature extraction unit is in communication connection with the input processing module; the dual-channel feature extraction unit constitutes a dual-channel attention mechanism, including a general semantic channel composed of a large language model and a domain feature channel, the general semantic channel receives the input text sent by the input processing module, extracts general language features through the Transformer layer of the LLM of the large language model; the domain feature channel communicates with the content construction unit to query whether the domain feature word is contained in the input text and the similarity with the domain feature word, and obtains the associated word content or similar associated word content; Interactive content generation module, the interactive content generation module analyzes and determines the interactive information corresponding to the input text according to the general language features and the domain features; the interactive information is content closed loop data composed of result data and analysis data; the analysis data is provided by the industry terminology library and the three-dimensional associated library; The correction module communicates with the interactive content generation module, determines that information deviation occurs after obtaining the result data and the analysis data, manually revises the result data or the analysis data through the correction module to obtain revised data; and the general language features and the domain feature words are used as retrieval items and the revised data is used as binding revised data and uploaded to the revised database; When only the result data or only the analysis data is modified, the unmodified data still maintains the initial state, and the modified data is separately marked; when the input text coincides with the retrieval items in the revised database, the revised data content is outputted preferentially; when the result data or the analysis data is modified, the modified data is updated and supplemented to the three-dimensional associated library synchronously; The analysis process of the interactive content generation module according to the general language features and the domain features is as follows: the large language model processes the input text to obtain feature data of general language features and domain features, determines explicit intent and implicit intent according to the feature data, the explicit intent is the intent that can be directly obtained through the feature data, the implicit intent is the emotional intent or the intent that cannot be directly obtained and needs to be obtained through reasoning, the reasoning process is as follows: each feature element in the general language features and the domain features is used as a reasoning logic neuron, wherein the feature element of the domain feature is the domain feature word and its own function; the large language model takes the reasoning logic neuron as a basic element, combines each basic element, and performs semantic reasoning to obtain N groups of reasoning combinations, and arranges the combinations from high to low by inducting the reasoning combinations, and selecting a reasoning result by guiding selection through the system active inquiry mode, and taking the result as the implicit intent; Finally, the large language model outputs result data matching the explicit or implicit intent, and analysis data corresponding to the structural data.

2. The large language model based content closed loop interaction system of claim 1, wherein: The result data consists of one or more results in a set, and the analysis data provides reasoning evidence for each subset in the set, the reasoning evidence being an intent, the intent having a reasoning relationship edge with a subset or an element of a subset in the three-dimensional association library, and the intent having a reasoning relationship edge with an industry standard in the industry term library.

3. The large language model based content closed loop interaction system of claim 1, wherein: When the interactive content generation module generates interactive information, the general language feature and the domain feature are obtained through the double-channel feature extraction unit, and the hit probability of the search term is calculated. When the domain feature of the search term coincides with the domain feature of the search term to reach a set value, it is considered as a hit; or the weight of each domain feature word of the search term is set, and when the weight value reaches a set value, it is considered as a hit; or the weight of the general language feature and the domain feature is set, and the coincidence amount of the general language feature and the domain feature of the input text and the search term is searched first, and then the total weight value is calculated. When the input text and the search term are considered as a hit, the interactive content generation module outputs the revision data content corresponding to the search term.

4. The large language model-based content closed-loop interaction system according to claim 3, characterized in that: The weight of the domain feature word of the search term is set as follows: after the correction module completes the revision of the revision data, the revision data binding process is entered, at which time the weight of the domain feature word of the search term can be set.

5. The large language model based content closed loop interaction system of claim 1, wherein: When the interactive content generation module generates interactive information, the general semantic and domain feature data are balanced through the gating mechanism.

6. The large language model based content closed loop interaction system of claim 1, wherein: After the analysis data is modified by the correction module, the modified data optimizes the large language model through the low-rank matrix incremental update method.

7. The large language model based content closed loop interaction system of claim 1, wherein: The double-channel feature extraction unit has a training module connected thereto, and the training module obtains a domain training set. The double-channel feature extraction unit obtains the domain training set, extracts the general language feature and the domain feature from the domain training set, and the large language model obtains the intent and the reasoning relationship edge corresponding to each data of the domain training set.

8. The large language model based content closed loop interaction system of claim 1, wherein: When the double-channel feature extraction unit obtains the domain feature word, the similar word is generated synchronously.

Images