An aviation equipment model named entity recognition method fusing multi-source knowledge

Abstract

This invention discloses a naming entity recognition method for aviation equipment models that integrates multi-source knowledge, comprising the following steps: S1: constructing a corpus for recognizing flat and nested entities of aviation equipment models; S2: establishing a naming entity recognition model integrating multi-source knowledge based on the corpus constructed in step S1; S3: using the naming entity recognition model established in step S2 to recognize the named entities of aviation equipment models. The naming entity recognition model integrating multi-source knowledge generally adopts an MRC model, integrating naming rule information into the query information in the input text, and combining RoBERTa and domain dictionary enhancement methods to extract semantic information during text encoding. The recognition method in this invention achieves F1 scores of 89.26% and 85.53% for flat entities and nested entities, respectively, realizing intelligent recognition of naming entities for aviation equipment models.

Description

Technical Field

[0001] This invention relates to the field of aviation knowledge engineering technology, and in particular to a method for naming and recognizing aviation equipment models by integrating multi-source knowledge. Background Technology

[0002] The development of aviation equipment models, as a pioneering step in aviation research, not only relates to the development of large-scale public transportation but also profoundly impacts the advancement of national strategic defense. The requirements analysis of aviation equipment models relies heavily on knowledge engineering or knowledge graphs. However, the data generated in actual engineering projects is often disorganized. The key to future equipment model development lies in extracting highly condensed knowledge from diverse and heterogeneous industry data.

[0003] Named entity recognition (NENT) serves as a bridge and link between natural language and professional knowledge transformation. It can not only achieve downstream tasks in traditional engineering, such as knowledge graph construction, but also provide knowledge-level decision-making for equipment model verification. Significant progress has been made in general-domain NENT tasks, primarily through dictionary-based methods, statistical machine learning methods, and deep learning methods. Dictionary-based methods heavily rely on manually pre-defined recognition templates and dictionaries. While this method ensures professionalism, dictionary and template construction is costly and has poor generalization ability. Statistical machine learning methods mainly include feature extraction and model building, with classic models such as Conditional Random Fields (CRF). However, statistical machine learning methods also require a large number of manually constructed features and perform poorly on small sample data. Deep learning methods have developed primarily from two dimensions: distributed word embedding representations and contextual neural network tables. The former is represented by the large-scale language pre-trained model BERT, while the latter mainly uses Bidirectional Long Short-Term Memory (BiLSTM) neural networks to capture contextual semantic relationships while CRF assists in label decoding.

[0004] Integrating domain knowledge into specific fields has always been a research challenge. From an engineering application perspective, traditional methods for academic research often extract tasks based on flat, single entities. However, flat, single entities cannot fully meet the information extraction needs of engineering applications, often resulting in nesting, ambiguity, and imbalance in engineering data. From the perspective of domain system knowledge, graph neural networks are commonly used. This involves embedding domain knowledge or domain knowledge graphs to form a distributed representation of the input, which is then fed into the model to identify and extract information. However, the biggest problem with graph neural networks lies in the complexity of their construction methods. Knowledge embedding relies on highly accurate information extraction, fusion, and classification models, which means a sharp increase in computational complexity and cost. From the perspective of the inherent knowledge information within words, a common approach is to incorporate the start and end character position information of words into the model based on Long Short-Term Memory (LSTM) neural networks, making it easier for the model to capture the boundary position information of entities. However, this method is only applicable to LSTM networks and is essentially ineffective in other sequence labeling models.

[0005] The identification of aircraft equipment models mainly faces the following two problems: 1) From a data perspective, mainstream named entity recognition tasks are based on supervised learning. Such methods heavily rely on a large amount of labeled data. The aviation data domain has distinct characteristics and even a certain degree of confidentiality. Publicly available data resources are extremely limited. Furthermore, the labeling process requires the guidance of professional knowledge engineering ontology modeling and the full integration of the dominant theoretical knowledge system in the domain. This means that entity recognition in the aviation domain is costly at the data level. 2) From a model perspective, named entity recognition methods are relatively mature in general domains. In particular, with the rise of the large-scale language pre-trained model BERT in recent years, the recognition accuracy is remarkable. However, in the aviation domain, general domain methods cannot effectively capture domain knowledge to complete entity recognition. Therefore, it is necessary to fully integrate professional knowledge into mainstream methods to improve recognition accuracy. Summary of the Invention

[0006] To address the aforementioned problems, this invention aims to provide a naming entity recognition method for aviation equipment models that integrates multi-source knowledge, thereby resolving the issues present in the prior art.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0008] A method for naming entities of aviation equipment models that integrates multi-source knowledge, characterized by comprising the following steps:

[0009] S1: Construct a corpus for flat and nested entity recognition of aviation equipment models;

[0010] S2: Based on the corpus constructed in step S1, establish a named entity recognition model that integrates multi-source knowledge;

[0011] S3: Use the named entity recognition model established in step S2 to identify the named entities of the aircraft equipment model.

[0012] Furthermore, the specific operation of step S1 includes the following steps:

[0013] S101: Collection and organization of multi-source heterogeneous aviation equipment model data;

[0014] S102: Annotate the collected corpus. The entity categories to be annotated include: aircraft, designers and manufacturers, relevant locations and countries and regions, aircraft engines, and aviation-related events.

[0015] S103: During the process of anticipating and labeling, a complete set of naming rules for aviation equipment models was summarized and formed.

[0016] Furthermore, the specific operation of step S101 includes the following steps:

[0017] S1011: Use web crawlers to obtain corpora related to aircraft equipment models from aviation websites;

[0018] S1012: Crawling internal data of aviation equipment from aviation professional websites and extracting aviation dictionaries and publicly available corpora;

[0019] S1013: General Corpus for Web Crawling.

[0020] Furthermore, the specific operation of step S2 includes the following steps:

[0021] S201: Using the corpus constructed in step S1 as the dataset, build a naming entity recognition framework that integrates aviation equipment model knowledge and input the corpus.

[0022] S202: The pre-trained model combining RoBERTa and wwm is used to extract weights and perform character-level encoding on the sentences to be recognized in the aviation equipment model named entity recognition task.

[0023] S203: Add the aviation dictionary extracted in step S1 to the vocabulary list for lexical-level encoding, and concatenate the character-level encoding result in step S202 with the lexical-level encoding result;

[0024] S204: Perform sequence matching on the sentences to be identified in the naming entity recognition task of aviation equipment models, solve the start and end indices, and obtain a naming entity recognition model that integrates multi-source knowledge;

[0025] S205: Train and test the named entity recognition model that integrates multi-source knowledge established in step S204.

[0026] Furthermore, the specific operation of step S201 includes the following steps:

[0027] S2011: Define the sentence to be recognized in the named entity recognition task of aviation equipment model as X = {x1, x2, ..., x}. N}, where x i This represents the i-th character in the sentence, where i = 1, 2, ..., N; N represents the maximum length of the sentence to be recognized.

[0028] S2012: The ultimate goal of the aviation equipment model name entity recognition task is defined as mapping each character to a specific category y, y∈Y, where Y represents the entity category, including aircraft, design and manufacturer, relevant location and country / region, aero-engine and aviation-related events;

[0029] S2013: Transform the machine reading comprehension task into a (Context, Query, Answer) triple dataset, where Context is the sentence to be recognized X = {x1, x2, ..., x...} N}, where Query is the question Q under a specific category y. y ={q1, q2, ..., q M}, where q j Let j represent the j-th character in the query, where j = 1, 2, ..., M; M is the maximum length of the query sentence; and Answer is the sequence of entities labeled with category Y, represented by the start and end positions x of the entities. start，end , start≤end; the final triplet (X, Q) is formed. y x start，end () corresponds to (Context, Query, Answer).

[0030] Furthermore, the specific operation of step S202 includes the following steps:

[0031] S2021: Query the corresponding Q y The sequence I = {[CLS], x1, x2, ..., x} is formed by concatenating the X corresponding to Context. N [SEP], q1, q2, ..., q M , [SEP]}, where [CLS] and [SEP] represent the start of the joint sequence and the sentence segmentation symbol;

[0032] S2022: Input the concatenated sequence I into the language pre-training model RoBERTa.

[0033]

[0034]

[0035] in, t represents the initial RoBERTa output weight vector. i W represents the i-th encoded embedded character. e and W b The parameters represent the learning parameters, 1≤l≤L, where L represents the total number of layers in RoBERTa and l represents each specific layer; Trm indicates that the Transformer block contains a multi-head attention layer, a fully connected layer, and a normalization layer. This represents the RoBERTa output weight vector of the (l-1)th layer; This represents the final weight vector obtained after RoBERTa-wwm; This represents the output weight vector of the Lth layer RoBERTa.

[0036] Furthermore, the specific operation of step S203 includes the following steps:

[0037] S2031: Add the aviation dictionary extracted in step S1 to the vocabulary for lexical-level encoding. For each given character x i and word list Construct sets for the beginning, middle, end, and single-word entities representing entity information.

[0038]

[0039]

[0040]

[0041]

[0042] Among them, w i，k This represents a subsequence {x} in the input text information I. i x i+1 , ..., x k};w j，k Represents the subsequence {x j x j+i , ..., x i x i+1 , ...x k};w j，i Represents the subsequence {x j x j+1 , ..., x i}, where i represents the different positions corresponding to the words, j represents the beginning position of the word, and k represents the end position of the word; B(x i) represents x i Corresponding to the vocabulary list The set of words at the beginning position, M(x) i ) represents x i Corresponding to the vocabulary list The set of words in the middle position, E(x) i ) represents x i Corresponding to the vocabulary list The set of words at the end position, S(x i ) represents a set of word entities;

[0043] S2032: Stable weights are formed by calculating word frequencies, wherein the vector weights of the starting characters in set B are calculated as follows:

[0044]

[0045] Among them, v s (B) represents the vector weight of the starting character; z(w) represents the word w in the vocabulary. The number of times it appears in ω word The weight vector obtained by word embedding method.

[0046] S2033: Repeat step S2032 to calculate v. s (M), v s (E) and v s (S);

[0047] S2034: The representations of the four vocabulary sets are fused into a fixed-dimensional feature to obtain the vocabulary-level encoding result, represented as follows:

[0048]

[0049] S2035: Convert the character-level encoding result obtained in step S202 into... Compared with the vocabulary-level encoding results obtained in step S2034 The data is concatenated to obtain the model's output.

[0050]

[0051]

[0052] in, E represents the output of the i-th character, and E represents the output of all characters.

[0053] Furthermore, the specific operation of step S204 includes the following steps:

[0054] S2041: Based on the model output obtained in step S203, calculate the probability distribution of the characters corresponding to the start and end indices under the Query condition.

[0055]

[0056]

[0057] Among them, L start and L end T represents the probability distribution of the characters corresponding to the start and end indices under the query condition. start and T end Learning weights in the process of solving for probability distributions;

[0058] S2042: Use argmax to solve for the start and end indices.

[0059]

[0060]

[0061] Where x and y represent the x-th and y-th rows of the matrix, given any start and end indices x. start ∈I start y end ∈I end ;

[0062] S2043: Use a binary classifier to train and predict the probability of start and end index matching.

[0063]

[0064] Where m represents the learning weights during the model training process.

[0065] Furthermore, the specific operation of step S205 includes the following steps:

[0066] S2051: Given a text X, the true category corresponding to the start and end indices of the text is represented by Y. start and Y end The loss functions for the start and end indices are calculated as Loss and Loss respectively. start and Loss end Y start，end This represents the category of the start and end index matching relationship for all queries, and its matching probability is represented by L. start，end The loss function is calculated as Loss. span ;

[0067] Loss start =CE(L start Y start )

[0068] Loss end =CE(L end Y end )

[0069] Loss span =CE(L start，end Y start，end );

[0070] S2052: Summing the three loss functions yields the training target loss function, Loss.

[0071] Loss = αLoss stan +βLoss end +γLoss span

[0072] Where α, β, γ∈[0,1] represent hyperparameters, and CE represents the cross-entropy loss function;

[0073] S2053: During the testing phase, based on I start and I end Select the start and end indices respectively, and output the final result based on the matching model.

[0074] The beneficial effects of this invention are:

[0075] 1. This invention combines the specialized nature of aviation-related corpora with the broad scope of general-domain corpora to construct a corpus for recognizing flat and nested entities in aviation equipment models. The corpus contains 377,670 characters, 8,063 sentences, 14,230 flat entities, and 5,681 nested entities. During the corpus acquisition, fusion, and annotation process, both an aviation-related dictionary and aviation equipment model naming rules were obtained. Based on the characteristics of the corpus, a named entity recognition model integrating multi-source knowledge was proposed. The model generally employs MRC (Machine Reading) and integrates naming rule information into the query information in the input text. During text encoding, RoBERTa and domain dictionary enhancement methods are combined to extract semantic information. Extensive experiments were conducted on the constructed dataset, achieving F1 scores of 89.26% and 85.53% for flat and nested entity recognition, respectively, thus realizing intelligent recognition of named entities for aviation equipment models.

[0076] 2. The named entity recognition model integrating multi-source knowledge established in this invention shows certain performance improvement over the BERT-MRC model on flat and nested entities, with F1 scores increasing by 0.82% and 2.38%, respectively. Quantitative analysis of the fused information shows that after removing prior knowledge of aircraft and engine naming rules from the query information, the performance of the flat and nested entity datasets decreases by 0.57% and 1.15%, respectively; after removing prior knowledge of domain expertise and entity boundary information from the dictionary augmentation method, the performance of the flat and nested entity datasets decreases by 0.69% and 1.50%, respectively. Ablation experiments demonstrate that the domain knowledge and entity boundary information provided by the dictionary augmentation method are more suitable for the entire recognition task. Attached Figure Description

[0077] Figure 1 This is a flowchart of the aviation equipment model naming entity recognition method that integrates multi-source knowledge according to the present invention.

[0078] Figure 2 This is a schematic diagram illustrating the corpus nesting phenomenon in Embodiment 1 of the present invention;

[0079] Figure 3 This is a schematic diagram illustrating an example of ambiguity in the corpus in Embodiment 1 of the present invention;

[0080] Figure 4 This is a schematic diagram illustrating the corpus imbalance phenomenon in Embodiment 1 of the present invention;

[0081] Figure 5 This is a structural diagram of the aviation equipment model name entity recognition model that integrates multi-source knowledge in this invention. Detailed Implementation

[0082] To enable those skilled in the art to better understand the technical solutions of the present invention, the technical solutions of the present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0083] As attached Figure 1 As shown, a method for naming entities of aviation equipment models that integrates multi-source knowledge includes the following steps:

[0084] S1: Construct a corpus for flat and nested entity recognition of aviation equipment models;

[0085] S2: Based on the corpus constructed in step S1, establish a named entity recognition model that integrates multi-source knowledge;

[0086] S3: Use the named entity recognition model established in step S2 to identify the named entities of the aircraft equipment model.

[0087] Specifically, step S1 includes the following steps:

[0088] S101: Collection and organization of multi-source heterogeneous aviation equipment model data;

[0089] The collection and organization of multi-source heterogeneous aviation equipment model corpora mainly needs to consider two factors: specialization and a certain scale. Specialization is an essential attribute of aviation equipment models, but the amount of data is small and extremely valuable, mainly from aviation-related websites and internal domain-specific data. For specialized data, this invention uses web crawlers to obtain corpora related to aviation equipment models from aviation-related websites, while simultaneously crawling internal aviation equipment data from aviation-specific websites and extracting aviation dictionaries and publicly available corpora. However, due to the scarcity of specialized corpora, it is insufficient to meet the training requirements of the model. Therefore, it is necessary to supplement with a large amount of publicly available general-purpose data. This invention combines the aviation dictionaries extracted from specialized corpora as an index list for web crawlers to obtain well-structured, large-scale general-purpose corpora such as Baidu Encyclopedia, Wikipedia, and Toutiao Encyclopedia to meet the scale requirements of model training.

[0090] S102: Annotate the collected corpus;

[0091] The core element of aircraft model naming entity recognition is the aircraft entity, but its derived attribute relationships are equally important, and these relationships can help extract more aircraft model information. Therefore, the corpus constructed in this invention requires annotation of five entity categories: aircraft, design and manufacturer, relevant locations and countries / regions, aircraft engines, and aviation-related events. The entire annotation process references professional aviation knowledge to formulate an annotation outline. In processing flat and nested named entities, flat entities are annotated based on the maximum length, while nested entities are annotated based on the smallest entity unit.

[0092] S103: During the process of anticipating and labeling, a complete set of naming rules for aviation equipment models was summarized and formed.

[0093] The knowledge fusion of the named entity recognition model mainly comes from the aviation-related dictionaries obtained by web crawling from professional websites and internal aviation data during corpus preparation. These dictionaries possess a specialization distinct from general domains and provide a reference for word frequency calculation in named entity recognition, thereby integrating knowledge information on the start, middle, and end points of entity boundaries and single-character entities. During corpus annotation, annotators summarized a complete set of naming rules for aviation equipment models from a large-scale corpus. Naming primarily uses model number and code, but can also combine multiple pieces of information such as country of origin, manufacturer, mixed Chinese and English numbers, codes, equipment category, and technical parameters. In summary, five key components are identified: 1) Model Number: The equipment model number, composed of English letters, hyphens, and numbers, is the core element of the entity name. Equipment models are unique within the same category, and generally, a single model number can directly refer to the equipment entity. 2) Code Number: The Chinese and English code number of the equipment entity, generally equivalent to the model number, can refer to the equipment entity independently. The same equipment may have different names in different countries; therefore, the code number and entity have a many-to-one relationship. 3) Category: In cases of ambiguous or conflicting model designations, equipment type is the best element for distinguishing entities. Additionally, special terms in the corpus mainly include "XX type," "XX level," and "XX model," with specific Chinese characters as suffixes. 4) Organization and Location: Primarily indicates the country or manufacturer of the equipment; both may coexist or exist singly. 5) Technical Parameters: Frequently found in engine entities and closely following the engine type, with horsepower being the primary representative. The rules are geared towards engineering applications in the aerospace field, incorporating prior knowledge to address issues such as abbreviation, nesting, ambiguity, and imbalance.

[0094] Step S2 includes the following steps:

[0095] S201: Using the corpus constructed in step S1 as the dataset, build a naming entity recognition framework that integrates aviation equipment model knowledge and input the corpus.

[0096] More specifically, S2011: Define the sentence to be recognized in the named entity recognition task for aviation equipment models as X = {x1, x2, ..., x}. N}, where x i This represents the i-th character in the sentence, where i = 1, 2, ..., N; N represents the maximum length of the sentence to be recognized.

[0097] S2012: The ultimate goal of the aviation equipment model name entity recognition task is defined as mapping each character to a specific category y, y∈Y, where Y represents the entity category, including aircraft, design and manufacturer, relevant location and country / region, aero-engine and aviation-related events;

[0098] S2013: Compared to traditional sequence labeling models, machine reading comprehension requires transforming the task into a (Context, Query, Answer) triple dataset, where Context is the sentence to be recognized X = {x1, x2, ..., x...}. N}, where Query is the question Q under a specific category y. y ={q1, q2, ..., q M}, where q j Let j represent the j-th character in the query, where j = 1, 2, ..., M; M is the maximum length of the query sentence; and Answer is the sequence of entities labeled with category Y, represented by the start and end positions x of the entities. start，end , start≤end; the final triplet (X, Q) is formed. y x start，end () corresponds to (Context, Query, Answer).

[0099] Furthermore, S202: The pre-trained model combining RoBERTa and wwm is used to extract weights and perform character-level encoding on the sentences to be recognized in the aviation equipment model named entity recognition task.

[0100] RoBERTa's biggest improvement lies in removing the prediction of the next sentence from the BERT model, allowing the model to focus more on extracting semantic relationships between words. The Wwm strategy, primarily proposed for Chinese corpora, offers the greatest improvement by adapting to Chinese word segmentation habits, restoring the Chinese word phrase form as much as possible during masking, rather than simply masking individual character predictions. The specific process includes...

[0101] S2021: Query the corresponding Q y The sequence I = {[CLS], x1, x2, ..., x} is formed by concatenating the X corresponding to Context. N [SEP], q1, q2, ..., q M , [SEP]}, where [CLS] and [SEP] represent the start of the joint sequence and the sentence segmentation symbol;

[0102] S2022: Input the concatenated sequence I into the language pre-training model RoBERTa.

[0103]

[0104]

[0105] in, t represents the initial RoBERTa output weight vector. i W represents the i-th encoded embedded character. e and Wb The parameters represent the learning parameters, 1≤l≤L, where L represents the total number of layers in RoBERTa and l represents each specific layer; Trm indicates that the Transformer block contains a multi-head attention layer, a fully connected layer, and a normalization layer. This represents the RoBERTa output weight vector of the (l-1)th layer; This represents the final weight vector obtained after RoBERTa-wwm; This represents the output weight vector of the Lth layer RoBERTa.

[0106] Furthermore, S203: Add the aviation dictionary extracted in step S1 to the vocabulary list for lexical-level encoding, and concatenate the character-level encoding result in step S202 with the lexical-level encoding result;

[0107] More specifically, S2031: Add the aviation dictionary extracted in step S1 to the vocabulary for lexical-level encoding. For each given character x i and word list Construct sets for the beginning, middle, end, and single-word entities representing entity information.

[0108]

[0109]

[0110]

[0111]

[0112] Among them, w i，k This represents a subsequence {x} in the input text information I. i x i+1 , ..., x k};w j，k Represents the subsequence {x j x j+i , ..., x i x i+1 , ...x k};w j，i Represents the subsequence {x j x j+1 , ..., x i}, where i represents the different positions corresponding to the words, j represents the beginning position of the word, and k represents the end position of the word; B(x i ) represents x i Corresponding to the vocabulary list The set of words at the beginning position, M(x) i ) represents x i Corresponding to the vocabulary list The set of words in the middle position, E(x) i ) represents x i Corresponding to the vocabulary list The set of words at the end position, S(x i ) represents a set of word entities;

[0113] S2032: Stable weights are formed by calculating word frequencies, wherein the vector weights of the starting characters in set B are calculated as follows:

[0114]

[0115] Among them, v s (B) represents the vector weight of the starting character; z(w) represents the word w in the vocabulary. The number of times it appears in ω word The weight vector obtained by word embedding method.

[0116] S2033: Repeat step S2032 to calculate v. s (M), v s (E) and v s (S);

[0117] S2034: The representations of the four vocabulary sets are fused into a fixed-dimensional feature to obtain the vocabulary-level encoding result, represented as follows:

[0118]

[0119] S2035: Convert the character-level encoding result obtained in step S202 into... Compared with the vocabulary-level encoding results obtained in step S2034 The data is concatenated to obtain the model's output.

[0120]

[0121]

[0122] in, E represents the output of the i-th character, and E represents the output of all characters.

[0123] Furthermore, S204: Perform sequence matching on the sentences to be identified in the aircraft equipment model named entity recognition task, and solve for the start and end indices;

[0124] S2041: Based on the model output obtained in step S203, calculate the probability distribution of the characters corresponding to the start and end indices under the Query condition.

[0125]

[0126]

[0127] Among them, L start and L end T represents the probability distribution of the characters corresponding to the start and end indices under the query condition. start and T end Indicates the weight of learning;

[0128] S2042: In text X, there may be multiple entities of the same category. This means that multiple start indices correspond to multiple end index values. For general flat entities, the start index corresponds to its nearest adjacent end index. However, this rule does not apply to nested entities. Therefore, argmax is used to solve for the start and end indices.

[0129]

[0130]

[0131] Where x and y represent the x-th and y-th rows of the matrix, given any start and end indices x. start ∈I start y end ∈I end ;

[0132] S2043: Use a binary classifier to train and predict the probability of start and end index matching.

[0133]

[0134] Where m represents the learning weights during the model training process.

[0135] The final named entity recognition model that integrates multi-source knowledge is shown in the attached figure. Figure 5 As shown.

[0136] Furthermore, S205: Train and test the model.

[0137] More specifically, S2051: Given a text context X, the true category representation corresponding to the start and end indices is Y. start and Y end The loss functions for the start and end indices are calculated as Loss and Loss respectively. start and Loss end Y start，end This represents the category of the start and end index matching relationship for all queries, and its matching probability is represented by L. start，end The loss function is calculated as Loss. span ;

[0138] Loss start=CE(L start Y start )

[0139] Loss end =CE(L end Y end )

[0140] Loss span =CE(L start，end Y start，end );

[0141] S2052: Summing the three loss functions yields the training target loss function, Loss.

[0142] Loss = αLoss stan +βLoss end +γLoss span

[0143] Where α, β, and Σ∈[0,1] represent hyperparameters, and CE represents the cross-entropy loss function;

[0144] S2053: During the testing phase, based on I start and I end Select the start and end indices respectively, and output the final result based on the matching model.

[0145] Example 1:

[0146] In this embodiment, a corpus is established using the method in step S1. During corpus selection, specialized aviation websites, industry-specific databases, and relatively well-structured general encyclopedic data are chosen to expand the training and test set. An aviation equipment model dictionary is used as the crawler's access list to help crawl the required encyclopedic corpus. The specialized and general corpora are then merged to form the original corpus, which undergoes preprocessing including duplicate data removal, special symbol deletion, and format conversion.

[0147] Analysis of the characteristics of the original corpus revealed the following three main features:

[0148] 1) Nested phenomenon

[0149] The corpus of aviation equipment model names contains a large number of nested named entities. Compared with traditional flat entities, nested entities can alleviate the phenomenon of abbreviation and reduction in entities. Furthermore, nested entities may originate from two or more categories, with different categories containing more information about aviation equipment models. (See attached...) Figure 2As shown, given the text to be identified, “F-117A is a stealth fighter developed by Lockheed Martin, Inc. of the United States”, “F-117A” corresponds to the aircraft category entity, and “Lockheed Martin, Inc. of the United States” corresponds to the manufacturing organization category entity. However, according to the nested entity rule, “United States” is a location category entity, and “Lockheed” is also a manufacturing organization category entity.

[0150] 2) Ambiguity

[0151] Because polysemy and polysemy are frequently encountered in corpus processing, this phenomenon, when it appears in aviation R&D terminology, can have serious consequences. For example, see attached... Figure 3 As shown, the given text "J-8 fighter jet, also known as J-8, nicknamed 'Airborne Handsome Man,' and NATO codenamed 'Finger Whale,'" also corresponds to the semantic meaning that "J-8 fighter jet" can be abbreviated as "J-8" or "J-8," nicknamed "Airborne Handsome Man," and NATO codenamed "Finger Whale."

[0152] 3) Imbalance phenomenon

[0153] Naming conventions vary significantly across different entity types. This leads to a large disparity in the number of entity categories (the long-tail effect) and also results in differences in the number of characters per entity type, as shown in the appendix. Figure 4 As shown, the given text “XB-29 uses four turbocharged Wright R-3350 twin-cyclone air-cooled radial engines.” contains approximately five aircraft entity characters, while the number of engine entity characters is more than three times that.

[0154] Based on the phenomena observed in the corpus, five entity types were identified: aircraft, aviation-related locations, aviation-related organizations, aircraft engines, and aviation-related events. These were then manually labeled, and the labeling results are shown in Table 1 below, which lists the labeled entities and their quantities.

[0155] Table 1 lists the types of entities and their quantities.

[0156]

[0157] Furthermore, the model is built and validated using the method in step S2. Specifically,

[0158] The aforementioned aviation equipment model dataset was divided into training, validation, and testing datasets in an 8:1:1 ratio. Based on the traditional annotation pattern BMES, the dataset was transformed into a text-based (Context, Query, Answer) data format, resulting in a total of 377,670 characters and 8,063 sentences. The input to the constructed named entity recognition model is shown in the attached figure. Figure 5As shown, the framework uses a machine reading comprehension named entity recognition framework as the baseline model. It incorporates aviation equipment model naming rules into the query information, integrates an aviation domain dictionary into the lexical embedding part, and fuses the information embedding using dictionary enhancement. Precision (P), recall (R), and F1 score are selected as evaluation metrics for entity recognition.

[0159] Table 2 below shows the results of named entity recognition for flat-panel aircraft equipment models. As can be seen from Table 2, the model proposed in this invention performs best compared to other models, achieving an F1 score of 89.26%. Compared to other models without using large-scale language pre-training models, CNN-CRF, BiLSTM-CRF, and Lattice-LSTM all generally meet the requirements of the named entity recognition task. Notably, the precision of the Lattice-LSTM model even surpasses that of the model proposed in this invention, and its recall and F1 score are also the best among the three models. Under the condition of using large-scale language pre-training models, the traditional BERT-BiLSTM-CRF model does not show significant fluctuations in recognition performance, while the method proposed in this invention shows a certain improvement compared to the BERT-MRC model, with an F1 score improvement of 0.82%.

[0160] Table 2 Comparison of Experimental Results for Flat Solids

[0161]

[0162] Table 3 below shows the nested aircraft equipment model name entity recognition results. In nested entity recognition, which is more in line with the target of extracting aircraft equipment model information, the model proposed in this invention has a more significant effect, with an F1 value of 85.53%. In comparison, the classic method Layered-BiLSTM-CRF is far inferior to the model proposed in this invention, and the F1 value is also improved by 2.38% compared with the BERT-MRC model.

[0163] Table 3 Comparison of Nested Entity Experiment Results

[0164]

[0165] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. An aircraft equipment model named entity recognition method fusing multi-source knowledge, characterized in that, Includes the following steps, S1: Construct a corpus for flat and nested entity recognition of aviation equipment models; S2: Based on the corpus constructed in step S1, establish a named entity recognition model that integrates multi-source knowledge; S3: Use the named entity recognition model established in step S2 to identify the named entities of the aviation equipment model. The specific operation of step S2 includes the following steps: S201: Using the corpus constructed in step S1 as the dataset, build a naming entity recognition framework that integrates aviation equipment model knowledge and input the corpus. S202: The pre-trained model combining RoBERTa and wwm is used to extract weights and perform character-level encoding on the sentences to be recognized in the aviation equipment model named entity recognition task. S203: Add the aviation dictionary extracted in step S1 to the vocabulary list for lexical-level encoding, and concatenate the character-level encoding result in step S202 with the lexical-level encoding result; S204: Perform sequence matching on the sentences to be identified in the naming entity recognition task of aviation equipment models, solve the start and end indices, and obtain a naming entity recognition model that integrates multi-source knowledge; S205: Train and test the named entity recognition model that integrates multi-source knowledge established in step S204; The specific operation of step S201 includes the following steps: S2011: define a to-be-recognized sentence in an aviation equipment model naming entity recognition task as wherein, denotes the i-th character in the sentence, i = 1, 2, …, N; N denotes the maximum length of the to-be-recognized sentence. S2012: The ultimate goal of the aircraft equipment model named entity recognition task is defined as mapping each character to a specific category y. Y represents the entity category, including aircraft, designers and manufacturers, relevant locations and countries and regions, aircraft engines, and aviation-related events; S2013: Transform the machine reading comprehension task into a (Context, Query, Answer) triple dataset, where Context is the sentence to be recognized. The query is a question under a specific category y. , where q j This represents the j-th character in the query, where j = 1, 2, ..., M; M is the maximum length of the query sentence; and Answer is the sequence of entities labeled with category Y, representing the start and end positions of the entities. , The final triplet formed Corresponding to (Context, Query, Answer); The specific operation of step S202 includes the following steps: S2021: Query corresponding to The sequence is formed by concatenating the X corresponding to the Context. ,in, and Indicates the beginning of a joint sequence and sentence delimiter; S2022: Input the concatenated sequence I into the language pre-training model RoBERTa. ； ； in, This represents the initial RoBERTa output weight vector. This represents the i-th encoded embedded character. and The parameters representing the learning process, L represents the total number of layers in RoBERTa, and l represents each specific layer; This indicates that the Transformer block contains a multi-head attention layer, a fully connected layer, and a normalization layer; Indicates the first The RoBERTa layer outputs a weight vector; This represents the final weight vector obtained after RoBERTa-wwm; This represents the RoBERTa output weight vector of the Lth layer; The specific operation of step S203 includes the following steps: S2031: Add the aviation dictionary extracted in step S1 to the vocabulary for lexical-level encoding. For each given character... and word list Construct sets for the beginning, middle, end, and single-word entities representing entity information. ； ； ； ； in, This represents a subsequence in the input text information I. ; Representing a subsequence ; Representing a subsequence , Indicates the different positions corresponding to words. Indicates the beginning position of a word. Indicates the end of a word; express Corresponding to the vocabulary list The set of words at the beginning position. express Corresponding to the vocabulary list The set of words in the middle position. express Corresponding to the vocabulary list The set of words at the end of a sentence. Represents a set of word entities; S2032: Stable weights are formed by calculating word frequencies, wherein the vector weights of the starting characters in set B are calculated as follows: ； in, The vector weight representing the start character; Indicates the word 'w' in the vocabulary. The number of times it appears in The weight vector obtained by word embedding method. ; S2033: Repeat step S2032, and calculate respectively. , and ; S2034: The representations of the four vocabulary sets are fused into a fixed-dimensional feature to obtain the vocabulary-level encoding result, represented as follows: ； S2035: Convert the character-level encoding result obtained in step S202 into... Compared with the vocabulary-level encoding results obtained in step S2034 The data is concatenated to obtain the model's output. ； ； in, E represents the output of the i-th character, and E represents the output of all characters.

2. The method for naming entities of aviation equipment models based on multi-source knowledge as described in claim 1, characterized in that, Step S1 includes the following steps: S101: Collection and organization of multi-source heterogeneous aviation equipment model data; S102: Annotate the collected corpus. The entity categories to be annotated include: aircraft, designers and manufacturers, relevant locations and countries and regions, aircraft engines, and aviation-related events. S103: During the process of anticipating and labeling, a complete set of naming rules for aviation equipment models was summarized and formed.

3. The method for naming entities of aviation equipment models based on multi-source knowledge as described in claim 2, characterized in that, The specific operation of step S101 includes the following steps: S1011: Use web crawlers to obtain corpora related to aircraft equipment models from aviation websites; S1012: Crawling internal data of aviation equipment from aviation professional websites and extracting aviation dictionaries and publicly available corpora; S1013: General Corpus for Web Crawling.

4. The method for naming entities of aviation equipment models based on multi-source knowledge as described in claim 1, characterized in that, The specific operation of step S204 includes the following steps: S2041: Based on the model output obtained in step S203, calculate the probability distribution of the characters corresponding to the start and end indices under the Query condition. ； ； in, and This represents the probability distribution of the characters corresponding to the start and end indices under the query condition. and Learning weights in the process of solving for probability distributions; S2042: Use argmax to solve for the start and end indices. ； ； Where x and y represent the x-th and y-th rows of the matrix, given arbitrary start and end indices. , ; S2043: Use a binary classifier to train and predict the probability of start and end index matching. ； Where m represents the learning weights during the model training process.

5. The method for naming entities of aviation equipment models based on multi-source knowledge according to claim 4, characterized in that, The specific operation of step S205 includes the following steps: S2051: Given the text context X, the true categories corresponding to the start and end indices are represented as follows: and The loss functions for the start and end indices are calculated as follows: and , This represents the category of the start and end index matching relationship for all queries, and its matching probability is expressed as... The loss function is calculated as follows: ; ； ； ； S2052: Summing the three loss functions yields the training target loss function. ; ； in, Denotes the hyperparameters, and CE represents the cross-entropy loss function; S2053: During the testing phase, based on and Select the start and end indices respectively, and output the final result based on the matching model.

Images