Natural language information extraction method and device for aviation support task

Abstract

The application relates to the technical field of natural language processing, in particular to a natural language information extraction method and device for an aviation support task, wherein the method comprises the following steps: obtaining to-be-reasoned natural language input data in the aviation support task; outputting a word initial score by using a small language model with a parameter quantity less than a preset parameter threshold; generating a structured candidate set based on a preset temperature coefficient and a parallel reasoning framework; extracting a structured candidate sample meeting a preset rule checking condition in the structured candidate set to construct a feasible solution set; and generating an actual reasoning result based on the feasible solution set. Thus, the defects of high algorithm power, long time delay, poor compatibility of self-defined extraction format, rule deviation caused by natural language ambiguity, weak generalization of template rules and weak supervision data set containing noise in related technologies are solved, and the problems of low aviation support colloquial text extraction precision, weak scene adaptation and unbalanced resource allocation are improved.

Description

Technical Field

[0001] This application relates to the field of natural language processing technology, and in particular to a method and apparatus for extracting natural language information for aviation support missions. Background Technology

[0002] In related technologies, a unified intermediate representation of documents can be generated by relying on visual language models and text parsing. The document complexity can be evaluated in three dimensions and three types of processing flows can be adaptively matched to output simplified intermediate format data. The extraction rules can be transformed by compiling prompt words with the help of extraction agents, and structured documents can be output after multiple verifications. Alternatively, a weakly supervised dataset in the aviation field can be built, and template rules and machine reading comprehension dual extraction models can be trained separately. The output of the two can be combined to complete the extraction of aviation document information.

[0003] However, among the related technologies, the multi-model serial architecture consumes a lot of computing power and has a large inference latency. The custom simple extraction intermediate format has insufficient compatibility and is easily affected by the ambiguity of natural language, resulting in rule deviation. Furthermore, the dataset constructed by weak supervision has annotation noise and poor ability to adapt template rules to new documents. These technologies cannot solve the core pain points of low accuracy, poor scenario adaptability, and unreasonable resource consumption in the structured extraction of colloquial text in aviation safety tasks, and urgently need improvement. Summary of the Invention

[0004] This application provides a method and apparatus for natural language information extraction for aviation support tasks, in order to solve the core pain points of related technologies, such as high computational consumption and large inference latency of multi-model serial architecture, insufficient compatibility of custom simple extraction intermediate format, susceptibility to rule deviation caused by natural language ambiguity, and the existence of annotation noise in weakly supervised datasets, poor ability of template rules to adapt to new documents, and inability to solve the core pain points of low accuracy of structured extraction of colloquial text in aviation support tasks, poor scenario adaptability, and unreasonable resource consumption.

[0005] The first aspect of this application provides a method for extracting natural language information for aviation support tasks, comprising the following steps: acquiring natural language input data to be inferred in the aviation support task; inputting the natural language input data to be inferred into a small language model with fewer than a preset parameter threshold, and outputting initial lexical scores; generating a structured candidate set matching the initial lexical scores based on a preset temperature coefficient and a parallel inference framework; extracting structured candidate samples that meet preset rule verification conditions from the structured candidate set, and constructing a feasible solution set based on the structured candidate samples; and generating the actual inference result of the natural language input data to be inferred based on the feasible solution set.

[0006] Optionally, in one embodiment of this application, the step of extracting structured candidate samples that satisfy preset rule verification conditions from the structured candidate set includes: matching the corresponding current scene information according to the natural language input data to be inferred; determining the corresponding domain constraint rules based on the current scene information; generating a scene-aware filtering unit based on the current scene information and the domain constraint rules; and filtering the structured candidate set through the scene-aware filtering unit to obtain structured candidate samples that satisfy the preset rule verification conditions.

[0007] Optionally, in one embodiment of this application, generating the actual inference result of the natural language input data to be inferred based on the feasible solution set includes: obtaining the model deployment configuration of the small language model; determining candidate aggregation strategies suitable for the small language model based on the model deployment configuration; selecting a target aggregation strategy from the candidate aggregation strategies based on the feasible solution set; and obtaining the actual inference result from the feasible solution set according to the target aggregation strategy.

[0008] Optionally, in one embodiment of this application, selecting a target aggregation strategy from the candidate aggregation strategies based on the feasible solution set includes: obtaining the occurrence frequency of structured candidate samples in the feasible solution set; and determining the target aggregation strategy based on the occurrence frequency.

[0009] Optionally, in one embodiment of this application, generating the actual reasoning result of the natural language input data to be reasoned based on the feasible solution set includes: if there is at least one structured candidate sample in the feasible solution set, then determining the actual reasoning result based on the feasible solution set; if there is no structured candidate sample in the feasible solution set, then inputting the natural language input data to be reasoned into a large language model, and using the large language model to output the reasoning result of the natural language input data to be reasoned.

[0010] A second aspect of this application provides a natural language information extraction device for aviation support tasks, comprising: an acquisition module for acquiring natural language input data to be inferred in the aviation support task; an output module for inputting the natural language input data to be inferred into a small language model with fewer than a preset parameter threshold and outputting initial word scores; a first generation module for generating a structured candidate set matching the initial word scores based on a preset temperature coefficient and a parallel inference framework; a construction module for extracting structured candidate samples that meet preset rule verification conditions from the structured candidate set and constructing a feasible solution set based on the structured candidate samples; and a second generation module for generating the actual inference result of the natural language input data to be inferred based on the feasible solution set.

[0011] Optionally, in one embodiment of this application, the construction module includes: a matching unit, configured to match corresponding current scene information based on the natural language input data to be inferred; a first determining unit, configured to determine corresponding domain constraint rules based on the current scene information; and a filtering unit, configured to generate a scene-aware filtering unit based on the current scene information and the domain constraint rules, and filter the structured candidate set through the scene-aware filtering unit to obtain structured candidate samples that meet the preset rule verification conditions.

[0012] Optionally, in one embodiment of this application, the second generation module includes: a first acquisition unit, configured to acquire the model deployment configuration of the small language model; a second determination unit, configured to determine a candidate aggregation strategy suitable for the small language model based on the model deployment configuration; and a second acquisition unit, configured to select a target aggregation strategy from the candidate aggregation strategies based on the feasible solution set, and acquire the actual inference result from the feasible solution set according to the target aggregation strategy.

[0013] Optionally, in one embodiment of this application, the second acquisition unit includes: an acquisition subunit, configured to acquire the occurrence frequency of structured candidate samples in the feasible solution set; and a determination subunit, configured to determine the target aggregation strategy based on the occurrence frequency.

[0014] Optionally, in one embodiment of this application, the second generation module includes: a third determining unit, configured to determine the actual reasoning result based on the feasible solution set when at least one structured candidate sample exists in the feasible solution set; and an input unit, configured to input the natural language input data to be reasoned into a large language model when the structured candidate sample does not exist in the feasible solution set, and output the reasoning result of the natural language input data to be reasoned using the large language model.

[0015] A third aspect of this application provides an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the natural language information extraction method for aviation support tasks as described in the above embodiments.

[0016] A fourth aspect of this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described natural language information extraction method for aviation support tasks.

[0017] A fifth aspect of this application provides a computer program product, including a computer program that, when executed, implements the above-described method for extracting natural language information for aviation support tasks.

[0018] This application embodiment can input natural language input data to be inferred in aviation support tasks into a small language model with fewer than a preset parameter threshold, thereby outputting initial word scores. Based on a preset temperature coefficient and a parallel inference framework, a matching structured candidate set is generated. Then, structured candidate samples that meet preset rule verification conditions are extracted from the structured candidate set, and a feasible solution set is constructed. Based on the feasible solution set, the corresponding actual inference result is determined. The small language model reduces the computational cost of inference, the parallel framework improves the generation efficiency, and rule filtering removes invalid candidates to reduce redundant calculations, ensuring that the inference output is compliant and accurate. Thus, it solves the core pain points of high computational cost and large inference latency of multi-model serial architecture, insufficient compatibility of custom simple extraction intermediate format, susceptibility to rule deviation caused by natural language ambiguity, and the existence of annotation noise in weakly supervised datasets, poor ability of template rules to adapt to new documents, and inability to solve the core pain points of low accuracy of structured extraction of colloquial text in aviation support tasks, poor scenario adaptability, and unreasonable resource consumption.

[0019] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description

[0020] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 This is a flowchart illustrating a natural language information extraction method for aviation support tasks, provided according to an embodiment of this application. Figure 2 A flowchart illustrating the working principle of a natural language information extraction method for aviation support tasks according to an embodiment of this application; Figure 3 This is a block diagram of a natural language information extraction device for aviation support missions provided according to an embodiment of this application; Figure 4 This is a schematic diagram of the structure of an electronic device provided according to an embodiment of this application. Detailed Implementation

[0021] The embodiments of this application are described in detail below. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application.

[0022] Natural Language Processing (NLP), a core subfield of artificial intelligence, enables machines to understand, interpret, and generate human language in a semantically meaningful way. It has evolved from early rule-based language parsing to become an indispensable component of modern AI, widely applied in machine translation, text analysis, and various other scenarios. Structured Natural Language Processing (SLP), a key sub-field of NLP, aims to transform unstructured text into formatted, directly usable structured information, thereby improving the efficiency of information retrieval, analysis, and subsequent applications. This technology has already seen initial applications in finance, healthcare, and other fields. Its implementation methods have undergone iterative upgrades from traditional statistical methods to deep learning and Transformer architectures, and currently, a development pattern has emerged where multiple technical paths coexist, including rule-based approaches, machine learning, and pre-trained language models.

[0023] In the field of aviation support missions, aviation support work encompasses multiple aspects such as flight scheduling, equipment maintenance, material supply, and personnel deployment. During actual execution, a large amount of conversational natural language text is generated, such as verbal instructions from dispatchers and task reports from frontline support personnel. Conversational texts are characterized by arbitrary sentence structures, abundant redundant information, a mixture of professional terminology and colloquial expressions, and strong contextual dependence. However, they contain core information about aviation support missions (such as mission type, responsible party, and supported objects). Structuring this type of conversational text and extracting key information is a prerequisite for achieving efficient scheduling, process control, accountability, and data-driven management of aviation support missions, and is of great significance for improving the response speed and execution accuracy of aviation support work.

[0024] Among related technologies, natural language structured processing technology has formed a certain technical foundation. With the increasing demand for data-driven and intelligent aviation support missions, the application of natural language structured processing technology in aviation support scenarios has gradually attracted attention, providing good technical support and exploration directions for the structured processing of spoken text in aviation support missions.

[0025] The following describes a method and apparatus for extracting natural language information for aviation support tasks, based on embodiments of the present application, with reference to the accompanying drawings. To address the core pain points mentioned in the background technology, such as high computational consumption and large inference latency of multi-model cascade architecture, insufficient compatibility of custom simplified extraction intermediate formats, susceptibility to rule deviations caused by natural language ambiguity, and the presence of annotation noise in weakly supervised datasets, poor adaptability of template rules to new documents, and inability to solve the problems of low accuracy, poor scenario adaptability, and unreasonable resource consumption in the structured extraction of colloquial text in aviation support tasks, this application provides a natural language information extraction method for aviation support tasks. In this method, the natural language input data to be inferred in aviation support tasks can be input into a small language model with fewer parameters than a preset parameter threshold, thereby outputting initial word scores. Based on a preset temperature coefficient and a parallel inference framework, a matching structured candidate set is generated. Then, structured candidate samples that meet the preset rule verification conditions are extracted from the structured candidate set, and feasible solution sets are constructed. Based on the feasible solution set, the corresponding actual inference result is determined. The small language model reduces the computational consumption of inference, the parallel framework improves the generation efficiency, and rule filtering removes invalid candidates to reduce redundant calculations, ensuring that the inference output is compliant and accurate. This addresses the core pain points in related technologies, such as high computational consumption and large inference latency in multi-model serial architectures, insufficient compatibility of custom simple extraction intermediate formats, susceptibility to rule deviations caused by natural language ambiguity, and the presence of annotation noise in datasets constructed with weak supervision, poor ability of template rules to adapt to new documents, and inability to solve the core pain points of low accuracy, poor scenario adaptability, and unreasonable resource consumption in the structured extraction of colloquial text in aviation safety tasks.

[0026] Specifically, Figure 1 This is a flowchart of a natural language information extraction method for aviation support tasks provided according to an embodiment of this application.

[0027] like Figure 1 As shown, this natural language information extraction method for aviation support tasks includes the following steps: In step S101, the natural language input data to be inferred in the aviation support mission is obtained.

[0028] It is understood that, in the embodiments of this application, the natural language input data to be inferred can be understood as the original natural language text used by large language models, small language models, etc. to perform text understanding, logical reasoning, question answering, and generation tasks. It may be, but is not limited to, derived from offline datasets, online real-time input, business structured conversion to natural language, etc.

[0029] Aviation support missions can be understood as a full-process supporting operation of aircraft and airborne equipment to ensure normal duty and operation. They cover all logistics and equipment support work, including military / civil aviation equipment maintenance, spare parts scheduling, airport operation and maintenance, fault repair, aviation material replenishment, pre-flight inspection, and equipment status support, and are used to limit the data source scenario.

[0030] In some embodiments, the present application can collect natural language input data to be processed in aviation support missions.

[0031] In step S102, the natural language input data to be inferred is input into a small language model with fewer than a preset parameter threshold, and the initial scores of the word units are output.

[0032] It is understood that, in the embodiments of this application, a small language model can be understood as a lightweight text model with fewer than a preset parameter threshold, characterized by low computational overhead, local offline deployment, and efficient processing of vertical tasks, distinguishing it from general-purpose large language models with hundreds of billions of parameters and cloud deployment. The preset parameter threshold can be set by those skilled in the art according to actual conditions, and this application does not impose specific limitations.

[0033] Furthermore, in this embodiment, the number of parameters can be understood as a core scale indicator of a small language model, referring to the total number of all neural network weight parameters within the model, typically measured in bytes (B). A larger number of parameters indicates a stronger ability to understand complex logic, but also higher computational consumption. For example, this embodiment selects a small language model with a parameter count between 4B and 12B.

[0034] The initial lexical score can be understood as the probability / weight score output by the small language model for each segmented lexical in the initial operation; the score represents the semantic importance and feature relevance of the lexical, and serves as the basic raw data for subsequent patent feature extraction and distinguishing feature comparison.

[0035] In some embodiments, the natural language input data to be inferred can be input into a small language model with fewer than a preset parameter threshold. The small language model can then be used to first split the natural language input data to be inferred into independent word units, calculate and output the initial word unit score for each word unit.

[0036] For example, based on the contradictory characteristics of sampling efficiency and model capability, this application embodiment adopts a small language model with a parameter size between 4B and 12B and deploys it on an edge device. Then, the natural language input data to be inferred is input into the small language model, and the small language model performs multiple independent decoding samplings on the input natural language input data to be inferred, and outputs the corresponding initial word scores.

[0037] In step S103, a structured candidate set matching the initial scores of the tokens is generated based on a preset temperature coefficient and a parallel reasoning framework.

[0038] It is understood that in this embodiment, the temperature coefficient, as a parameter of the small language model, can be used to control the diversity of the small language model's output; the higher the temperature coefficient, the more diverse the output. This embodiment ensures diversity by setting a preset temperature coefficient (e.g., a high temperature coefficient, such as a temperature coefficient > 0.8). That is, by utilizing the output diversity brought about by sampling with a high temperature coefficient, the probability of including at least one correct result in the structured candidate set is significantly increased, thereby compensating for the poor stability of single predictions in the small language model. The preset temperature coefficient can be set by those skilled in the art according to actual conditions, and this application does not impose specific limitations.

[0039] The parallel inference framework supports an accelerated inference architecture that simultaneously computes multiple text segments and multiple lexical units. This framework may include, but is not limited to, vLLM (Virtual Large Language Model). Furthermore, this embodiment selects the vLLM framework as the parallel inference framework, meaning it infers multiple input samples simultaneously, resulting in faster inference speed compared to directly using existing frameworks.

[0040] A structured candidate set can be understood as a dataset of compliant technical terms that have been screened and organized into a hierarchical and standardized format according to fixed rules.

[0041] In some embodiments, the present application may employ a preset temperature coefficient, combined with a parallel inference framework capable of batch synchronous computation, to generate an adapted structured candidate set based on the initial lexical scores.

[0042] For example, in this embodiment of the application, the preset temperature coefficient can be set to a high temperature coefficient (generally greater than 0.8) and multiple samplings can be performed to generate a batch (generally 8-16 visible models) of differentiated structured candidate sets. During the sampling process, vLLM is used as a parallel inference framework, and all sampling requests are executed synchronously through a batch parallel inference mechanism to reduce the increase in inference latency caused by the increase in the number of samples, thus keeping the additional time consumption caused by increasing the number of samplings within a small range.

[0043] In step S104, structured candidate samples that meet the preset rule verification conditions are extracted from the structured candidate set, and a feasible solution set is constructed based on the structured candidate samples.

[0044] It is understood that, in the embodiments of this application, the preset rule verification conditions can be understood as pre-configured filtering and judgment rules, which can be specifically set by those skilled in the art according to the actual situation, and this application does not impose specific limitations. For example, the embodiments of this application can construct preset rule verification conditions based on one or more of the following: scenario prior verification based on a predefined entity list, spatiotemporal logic verification based on standard operating procedures, mutual exclusion verification between tasks, and dependency verification between tasks. By injecting deterministic domain knowledge into the post-processing flow of the small language model in a pluggable manner, erroneous outputs generated by the model that do not conform to the constraints of the physical world can be effectively filtered.

[0045] Furthermore, in this embodiment of the application, the structured candidate set is filtered according to preset rule verification conditions to obtain structured candidate samples.

[0046] The feasible solution set is a complete technical solution combination library formed by combining and arranging all structured candidate samples that meet the preset rule verification conditions in the structured candidate set.

[0047] In some embodiments, the present application can perform verification filtering on the structured candidate set according to preset verification conditions, screen out structured candidate samples that meet the preset rule verification conditions, and construct a feasible solution set based on the structured candidate samples.

[0048] For example, embodiments of this application may construct preset rule verification conditions based on one or more of the following: scenario prior verification based on a predefined entity list, spatiotemporal logic verification based on standard operating procedures, inter-task mutual exclusion verification, and inter-task dependency verification.

[0049] Furthermore, in this embodiment, the structured candidate set is validated and filtered according to preset rule validation conditions. It determines whether each structured candidate sample in the set satisfies the preset rule validation conditions, and if so, retains the corresponding structured candidate sample and adds it to the feasible solution set. By utilizing deterministic external knowledge to strongly correct the uncertain output of the small language model, the space for incorrect answers is significantly reduced.

[0050] Optionally, in one embodiment of this application, extracting structured candidate samples that meet preset rule verification conditions from the structured candidate set includes: matching the corresponding current scene information based on the natural language input data to be inferred; determining the corresponding domain constraint rules based on the current scene information; generating a scene-aware filtering unit based on the current scene information and the domain constraint rules; and filtering the structured candidate set through the scene-aware filtering unit to obtain structured candidate samples that meet preset rule verification conditions.

[0051] It is understood that, in the embodiments of this application, the current scene information can be understood as the current scene context information of the aviation hub scene, which may include, but is not limited to, aircraft status, mission status, etc.

[0052] Domain constraint rules may include, but are not limited to, a legal list of aircraft and equipment within the airport, temporal dependencies defined in standard operating procedures, and mutual exclusion rules between tasks.

[0053] The scene-aware filtering unit can be understood as an independently callable logical filtering module that can selectively filter structured candidate samples that do not conform to domain constraint rules, and obtain structured candidate samples that meet the preset rule verification conditions, such as a perception filter.

[0054] Furthermore, in the embodiments of this application, the rule types of the scene-aware filtering unit may include, but are not limited to, format errors, entity errors, non-compliance with standard operating procedures, and unreasonable task status, etc., and this application does not impose specific limitations.

[0055] In this context, a formatting error can be understood as the extracted information lacking elements, such as missing aircraft entities, parking position entities, or mission status information.

[0056] Entity errors can be understood as the retrieved information not existing in the database, such as not finding the current aircraft or location, or the mission status name being non-existent or non-standard.

[0057] Non-compliance with standard operating procedures can be understood as starting a task before the prerequisite task that a certain task depends on has been completed.

[0058] An unreasonable task status can be understood as the extracted information being unreasonable in the current real-time state of the database, such as a task being completed before it has even started.

[0059] In some embodiments, the present application can identify the corresponding current scene information from the natural language input data to be reasoned, retrieve the corresponding domain constraint rules based on the current scene information, and then generate a scene-aware filtering unit by combining the current scene information and the domain constraint rules. The scene-aware filtering unit is used to filter the structured candidate set to obtain structured candidate samples that meet the preset rule verification conditions.

[0060] For example, embodiments of this application can construct a context-aware filter, i.e., a scene-aware filtering unit, that is, deeply bound to the aviation hub scenario. This perception filter filters the structured candidate set based on current scenario information (such as aircraft status, task status, etc.) using domain constraint rules, including but not limited to a legal list of aircraft and equipment within the airport, temporal dependencies defined in standard operating procedures, and mutual exclusion rules between tasks. For each structured candidate sample in the structured candidate set, it is input into the perception filter one by one. The perception filter verifies whether the structured candidate sample meets the domain constraints. If it does, it is determined that the preset rule verification condition is met, and the corresponding structured candidate sample is retained, completing the construction of the feasible solution set. This approach utilizes deterministic external knowledge to strongly correct the uncertain output of the language model, significantly reducing the space for incorrect answers.

[0061] In step S105, based on the feasible solution set, the actual reasoning result of the natural language input data to be reasoned is generated.

[0062] In some embodiments, the present application can determine the actual reasoning result corresponding to the natural language input data to be reasoned based on the feasible solution set.

[0063] Optionally, in one embodiment of this application, generating the actual reasoning result of the natural language input data to be reasoned based on the feasible solution set includes: if there is at least one structured candidate sample in the feasible solution set, then determining the actual reasoning result based on the feasible solution set; if there is no structured candidate sample in the feasible solution set, then inputting the natural language input data to be reasoned into the large language model, and using the large language model to output the reasoning result of the natural language input data to be reasoned.

[0064] In some embodiments, the present application can determine the actual reasoning result based on the feasible solution set when there is at least one structured candidate sample in the feasible solution set.

[0065] For example, in embodiments of this application, when the feasible solution set is not empty and there is at least one structured candidate sample, the corresponding actual reasoning result can be determined based on the feasible solution set and the small language model.

[0066] In some embodiments, when there are no structured candidate samples in the feasible solution set, the large language model can be used to process the natural language input data to be reasoned, so as to output the reasoning result of the natural language input data to be reasoned.

[0067] For example, this application embodiment designs a dynamic model routing mechanism as a fallback solution. When all structured candidate samples are determined to be invalid by the perception filter, the natural language input data to be inferred is determined to be a complex sentence. At this time, the routing action is automatically triggered, forwarding the natural language input data to be inferred and the current scene information to a larger language model with stronger computing power (such as a large language model with more than or equal to 30B parameters) for processing, thereby determining the corresponding inference result, or switching to manual intervention for confirmation. This ensures that in most normal cases, the small language model ensures low latency, and in a few difficult cases, the large language model is called to ensure high accuracy, thus achieving the optimal balance between processing efficiency and extraction accuracy overall.

[0068] The dynamic model routing mechanism proposed in this application can be understood as a key technical means to achieve a balance between system-level accuracy and efficiency. Specifically, it uses a small language model as the default processing channel and the condition of whether the feasible solution set is empty: only when all structured candidate samples generated by multiple samplings of the small language model are determined to be invalid by the perceptual filter, is the natural language input data to be inferred automatically transferred to the large language model or the manual processing stage. This allows for dynamic matching of computing resources with the complexity of the input samples, significantly reducing average processing latency while maintaining high average accuracy.

[0069] Optionally, in one embodiment of this application, generating the actual reasoning result of the natural language input data to be reasoned based on the feasible solution set includes: obtaining the model deployment configuration of the small language model; determining candidate aggregation strategies suitable for the small language model based on the model deployment configuration; selecting a target aggregation strategy from the candidate aggregation strategies based on the feasible solution set; and obtaining the actual reasoning result from the feasible solution set according to the target aggregation strategy.

[0070] It is understood that, in the embodiments of this application, the model deployment configuration may include, but is not limited to, hardware parameters and business configuration information required for model deployment and operation, such as model size, application focus, computing power, inference latency, context window, batch processing limit, and memory usage threshold.

[0071] Candidate aggregation strategies may include, but are not limited to, parameterless fixed aggregation, such as mean pooling, weighted length aggregation, and end-to-end splicing aggregation; learnable weighted aggregation, such as self-attention aggregation and gated pooling aggregation; and higher-order hybrid aggregation, such as mean and maximum splicing aggregation and hierarchical aggregation.

[0072] The target aggregation strategy can be understood as the unique and optimal merging scheme that is selected from all candidate aggregation strategies and is suitable for the current model deployment configuration.

[0073] In some embodiments, the present application can read the model deployment configuration of a small language model, and based on the model deployment configuration, filter out all candidate aggregation strategies that are adapted to the deployment environment, and then determine the target aggregation strategy according to the feasible solution set, and then use the target aggregation strategy to obtain the actual inference result from the feasible solution set.

[0074] For example, embodiments of this application can select different candidate aggregation strategies based on model deployment configurations such as model size and application focus, and then select weighted length aggregation as the target aggregation strategy. That is, the product of the frequency of the structured candidate sample appearing in the feasible solution set and the number of structured answer entries contained therein is used as the scoring criterion, and the one with the highest score is selected. This strategy can simultaneously take into account the model's confidence in the answer and the information completeness of the answer.

[0075] Furthermore, in this embodiment of the application, the actual reasoning results are obtained from the feasible solution set according to the target aggregation strategy.

[0076] Optionally, in one embodiment of this application, selecting a target aggregation strategy from candidate aggregation strategies based on a feasible solution set includes: obtaining the occurrence frequency of structured candidate samples in the feasible solution set; and determining the target aggregation strategy based on the occurrence frequency.

[0077] It is understood that, in the embodiments of this application, the frequency of occurrence can be understood as the ratio of the number of times the same structured candidate sample appears repeatedly in the feasible solution set to the total number of samples.

[0078] In some embodiments, the present application can traverse the structured candidate samples in the feasible solution set and count the occurrence frequency of each structured candidate sample, and then determine the target aggregation strategy based on the occurrence frequency.

[0079] The working principle of the natural language information extraction method for aviation support tasks proposed in the embodiments of this application will be introduced below.

[0080] Specifically, Figure 2 This is a flowchart illustrating the working principle of a natural language information extraction method for aviation support tasks provided according to an embodiment of this application.

[0081] Step S201: Candidate generation stage.

[0082] In this embodiment, a small language model with fewer than a preset parameter threshold, deployed on an edge device, is used to perform multiple independent decoding samples of the natural language input data to be inferred. By applying a high temperature coefficient (e.g., temperature coefficient > 0.8) and performing multiple samplings, a batch of differentiated structured candidate sets is generated. Furthermore, vLLM is used as a parallel inference framework during the sampling process, and all sampling requests are executed synchronously through a batch parallel inference mechanism.

[0083] Step S202: Consistency filtering process.

[0084] In this embodiment, a context-aware filter deeply bound to the aviation hub scenario can be constructed. Then, the perception filter is used to verify each structured candidate sample in the structured candidate set one by one, detect whether the structured candidate sample meets the domain constraints, and when it does, determine that the preset rule verification conditions are met, and retain the corresponding structured candidate sample, thereby completing the construction of the feasible solution set.

[0085] Step S203: Candidate aggregation and selection process.

[0086] In this embodiment, the actual inference result can be obtained from the target aggregation strategy in the feasible solution set. Specifically, different candidate aggregation strategies are selected based on the model deployment configuration such as model size and application focus. Then, based on the frequency of occurrence of structured candidate samples in the feasible solution set, weighted length aggregation is selected as the target aggregation strategy to determine the actual inference result.

[0087] Step S204: Dynamic model routing mechanism.

[0088] In this embodiment, when there are no structured candidate samples in the feasible solution set, the natural language input data to be inferred is determined to be a complex sentence. At this time, the routing action is automatically triggered to forward the natural language input data to be inferred and the current scene information to a large language model with stronger computing power, or to switch to manual intervention for confirmation processing, thereby outputting the corresponding inference result.

[0089] The natural language information extraction method for aviation support tasks proposed in this application can input the natural language input data to be inferred in the aviation support task into a small language model with fewer than a preset parameter threshold, thereby outputting initial word scores. Based on a preset temperature coefficient and a parallel inference framework, a matching structured candidate set is generated. Then, structured candidate samples that meet the preset rule verification conditions are extracted from the structured candidate set, and a feasible solution set is constructed. Based on the feasible solution set, the corresponding actual inference result is determined. The small language model reduces the computational cost of inference, the parallel framework improves the generation efficiency, and rule filtering removes invalid candidates to reduce redundant calculations, ensuring that the inference output is compliant and accurate. Thus, it solves the core pain points of high computational cost and large inference latency of multi-model serial architecture, insufficient compatibility of custom simple extraction intermediate format, susceptibility to rule deviation caused by natural language ambiguity, and the existence of annotation noise in weakly supervised datasets, poor ability of template rules to adapt to new documents, and inability to solve the core pain points of low accuracy of structured extraction of colloquial text in aviation support tasks, poor scenario adaptability, and unreasonable resource consumption.

[0090] Next, referring to the accompanying drawings, a natural language information extraction device for aviation support tasks is described according to an embodiment of this application.

[0091] Figure 3 This is a block diagram of a natural language information extraction device for aviation support missions provided according to an embodiment of this application.

[0092] like Figure 3 As shown, the natural language information extraction device 10 for aviation support missions includes: an acquisition module 100, an output module 200, a first generation module 300, a construction module 400, and a second generation module 500.

[0093] The acquisition module 100 is used to acquire the natural language input data to be inferred in the aviation support mission.

[0094] The output module 200 is used to input the natural language input data to be inferred into a small language model with fewer than the preset parameter threshold, and output the initial scores of the word units.

[0095] The first generation module 300 is used to generate a structured candidate set that matches the initial scores of word units based on a preset temperature coefficient and a parallel reasoning framework.

[0096] The construction module 400 is used to extract structured candidate samples that meet the preset rule verification conditions from the structured candidate set, and to construct a feasible solution set based on the structured candidate samples.

[0097] The second generation module 500 is used to generate the actual reasoning results of the natural language input data to be reasoned based on the feasible solution set.

[0098] Optionally, in one embodiment of this application, the construction module 400 includes: a matching unit, a first determining unit, and a filtering unit.

[0099] The matching unit is used to match the corresponding current scene information based on the natural language input data to be inferred.

[0100] The first determining unit is used to determine the corresponding domain constraint rules based on the current scene information.

[0101] The filtering unit is used to generate a scene-aware filtering unit based on the current scene information and domain constraint rules. The scene-aware filtering unit filters the structured candidate set to obtain structured candidate samples that meet the preset rule verification conditions.

[0102] Optionally, in one embodiment of this application, the second generation module 500 includes: a first acquisition unit, a second determination unit, and a second acquisition unit.

[0103] The first acquisition unit is used to acquire the model deployment configuration of the small language model.

[0104] The second determining unit is used to determine candidate aggregation strategies suitable for small language models based on the model deployment configuration.

[0105] The second acquisition unit is used to select a target aggregation strategy from the candidate aggregation strategies based on the feasible solution set, and to obtain the actual reasoning result from the feasible solution set according to the target aggregation strategy.

[0106] Optionally, in one embodiment of this application, the second acquisition unit includes: an acquisition subunit and a determination subunit.

[0107] The acquisition sub-unit is used to obtain the frequency of occurrence of structured candidate samples in the feasible solution set.

[0108] Determine sub-units to determine the target aggregation strategy based on occurrence frequency.

[0109] Optionally, in one embodiment of this application, the second generation module 500 includes: a third determining unit and an input unit.

[0110] The third determining unit is used to determine the actual reasoning result based on the feasible solution set when there is at least one structured candidate sample in the feasible solution set.

[0111] The input unit is used to input the natural language input data to be reasoned into the large language model when there are no structured candidate samples in the feasible solution set, and to output the reasoning result of the natural language input data to be reasoned using the large language model.

[0112] It should be noted that the explanation of the embodiment of the natural language information extraction method for aviation support missions also applies to the natural language information extraction device for aviation support missions in this embodiment, and will not be repeated here.

[0113] The natural language information extraction device for aviation support tasks proposed in this application can input the natural language input data to be inferred in aviation support tasks into a small language model with fewer than a preset parameter threshold, thereby outputting initial word scores. Based on a preset temperature coefficient and a parallel inference framework, a matching structured candidate set is generated. Then, structured candidate samples that meet preset rule verification conditions are extracted from the structured candidate set, and a feasible solution set is constructed. Based on the feasible solution set, the corresponding actual inference result is determined. The small language model reduces the computational power consumption of inference, the parallel framework improves the generation efficiency, and rule filtering removes invalid candidates to reduce redundant calculations, ensuring that the inference output is compliant and accurate. Thus, it solves the core pain points of high computational power consumption and large inference latency of multi-model serial architecture, insufficient compatibility of custom simple extraction intermediate format, susceptibility to rule deviation caused by natural language ambiguity, and the existence of annotation noise in weakly supervised datasets, poor ability of template rules to adapt to new documents, and inability to solve the core pain points of low accuracy of structured extraction of colloquial text in aviation support tasks, poor scenario adaptability, and unreasonable resource consumption.

[0114] Figure 4 This is a schematic diagram of the structure of an electronic device provided according to an embodiment of this application. The electronic device may include: The memory 401, the processor 402, and the computer program stored on the memory 401 and capable of running on the processor 402.

[0115] When the processor 402 executes the program, it implements the natural language information extraction method for aviation support tasks provided in the above embodiments.

[0116] Furthermore, electronic devices also include: Communication interface 403 is used for communication between memory 401 and processor 402.

[0117] The memory 401 is used to store computer programs that can run on the processor 402.

[0118] Memory 401 may include high-speed RAM memory, and may also include non-volatile memory, such as at least one disk storage device.

[0119] If the memory 401, processor 402, and communication interface 403 are implemented independently, then the communication interface 403, memory 401, and processor 402 can be interconnected via a bus to complete communication between them. The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized into address buses, data buses, control buses, etc. For ease of representation, Figure 4 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0120] Optionally, in a specific implementation, if the memory 401, processor 402, and communication interface 403 are integrated on a single chip, then the memory 401, processor 402, and communication interface 403 can communicate with each other through an internal interface.

[0121] Processor 402 may be a central processing unit (CPU), an application specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of this application.

[0122] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described natural language information extraction method for aviation support tasks.

[0123] This application also provides a computer program product, including a computer program that, when executed, implements the above-described method for extracting natural language information for aviation support tasks.

[0124] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0125] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "N" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0126] Any process or method described in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or N executable instructions for implementing custom logic functions or processes, and the scope of the preferred embodiments of this application includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as should be understood by those skilled in the art to which embodiments of this application pertain.

[0127] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of computer-readable media include: an electrical connection having one or more wires (electronic device), a portable computer disk drive (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Alternatively, the computer-readable medium may be paper or other suitable media on which the program can be printed, since the program can be obtained electronically by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in a computer memory.

[0128] It should be understood that the various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, the N steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. If implemented in hardware, as in another embodiment, it can be implemented using any one or more of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0129] Those skilled in the art will understand that all or part of the steps of the methods in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.

[0130] Furthermore, the functional units in the various embodiments of this application can be integrated into a processing module, or each unit can exist physically separately, or two or more units can be integrated into a module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium.

[0131] The storage medium mentioned above can be a read-only memory, a disk, or an optical disk, etc. Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of this application.

Claims

1. A method for extracting natural language information for aviation support tasks, characterized in that, Includes the following steps: Acquire natural language input data to be inferred in aviation support missions; The natural language input data to be inferred is input into a small language model with fewer than a preset parameter threshold, and the initial word score is output. Based on a preset temperature coefficient and a parallel reasoning framework, a structured candidate set matching the initial score of the word element is generated; From the structured candidate set, extract structured candidate samples that meet the preset rule verification conditions, and construct a feasible solution set based on the structured candidate samples; Based on the feasible solution set, the actual reasoning result of the natural language input data to be reasoned is generated.

2. The method according to claim 1, characterized in that, Extracting structured candidate samples that satisfy preset rule verification conditions from the structured candidate set includes: Match the corresponding current scene information based on the natural language input data to be inferred; Based on the current scenario information, determine the corresponding domain constraint rules; A scene-aware filtering unit is generated based on the current scene information and the domain constraint rules. The scene-aware filtering unit is used to filter the structured candidate set to obtain structured candidate samples that meet the preset rule verification conditions.

3. The method according to claim 1, characterized in that, The step of generating the actual reasoning result based on the feasible solution set for the natural language input data to be reasoned includes: Obtain the model deployment configuration of the small language model; Based on the model deployment configuration, determine the candidate aggregation strategies suitable for the small language model; Based on the feasible solution set, a target aggregation strategy is selected from the candidate aggregation strategies, and the actual reasoning result is obtained from the feasible solution set according to the target aggregation strategy.

4. The method according to claim 3, characterized in that, The step of selecting a target aggregation strategy from the candidate aggregation strategies based on the feasible solution set includes: Obtain the frequency of occurrence of structured candidate samples in the feasible solution set; The target aggregation strategy is determined based on the frequency of occurrence.

5. The method according to claim 1, characterized in that, The step of generating the actual reasoning result based on the feasible solution set for the natural language input data to be reasoned includes: If there is at least one structured candidate sample in the feasible solution set, the actual reasoning result is determined based on the feasible solution set. If the structured candidate sample does not exist in the feasible solution set, the natural language input data to be reasoned is input into the large language model, and the reasoning result of the natural language input data to be reasoned is output using the large language model.

6. A natural language information extraction device for aviation support missions, characterized in that, include: The acquisition module is used to acquire the natural language input data to be inferred in aviation support missions; The output module is used to input the natural language input data to be inferred into a small language model with fewer than a preset parameter threshold, and output the initial scores of the word units. The first generation module is used to generate a structured candidate set that matches the initial score of the word based on a preset temperature coefficient and a parallel inference framework. The construction module is used to extract structured candidate samples that meet the preset rule verification conditions from the structured candidate set, and construct a feasible solution set based on the structured candidate samples; The second generation module is used to generate the actual reasoning result of the natural language input data to be reasoned based on the feasible solution set.

7. The apparatus according to claim 6, characterized in that, The building module includes: The matching unit is used to match the corresponding current scene information based on the natural language input data to be inferred; The determining unit is used to determine the corresponding domain constraint rules based on the current scene information; The filtering unit is used to generate a scene-aware filtering unit based on the current scene information and the domain constraint rules, and to filter the structured candidate set through the scene-aware filtering unit to obtain structured candidate samples that meet the preset rule verification conditions.

8. An electronic device, characterized in that, include: A memory, a processor, and a computer program stored in the memory and executable on the processor, the processor executing the program to implement the natural language information extraction method for aviation support missions as described in any one of claims 1-5.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, The program is executed by the processor to implement the natural language information extraction method for aviation support tasks as described in any one of claims 1-5.

10. A computer program product, characterized in that, Includes a computer program, which, when executed, is used to implement the natural language information extraction method for aviation support missions as described in any one of claims 1-5.

Images