Ontology-driven structured extraction method for aviation intelligence table information

Abstract

The invention discloses an aviation intelligence table information structured extraction method based on ontology driving, and belongs to the technical field of aviation data information processing, and the method comprises the following steps: constructing an ontology framework model of aviation intelligence data; performing structured information extraction on the aviation intelligence data to obtain aviation intelligence table data; extracting English structured text data in the aviation intelligence table data through a regular expression in the regular expression matching template to obtain a structured entity; constructing a conditional random field (CRF) model, and performing named entity recognition on English semi-structured and non-structured text data in the aviation intelligence table data to obtain semi-structured and non-structured entities; and summarizing the structured entities, the semi-structured entities and the non-structured entities to obtain aviation intelligence data feature entities, and completing structured extraction of the aviation intelligence table information. According to the method, the problems that time and labor are wasted, errors are prone to occurring and automation is not enough when table information in data in the aviation intelligence field is manually searched at present are solved.

Description

Technical Field

[0001] The invention belongs to the technical field of aviation data information processing, and in particular relates to a method for extracting structured aviation intelligence table information based on ontology drive. Background Art

[0002] Aviation intelligence is an important source of information to support aviation operations. It contains massive, real-time data related to aviation operation safety, which lays an important foundation for the planning and real-time decision-making of aviation operations. At present, there is a large amount of aviation intelligence information in all countries in the world. Therefore, the automatic extraction of relevant aviation data can reduce a lot of manual processing time and support different downstream business scenarios.

[0003] At present, scholars at home and abroad mainly use methods based on templates and rules, machine learning, and deep learning in the research of named entity recognition (NER). The rule-based method has poor adaptability to complex texts and fields, and consumes a lot of manpower and material resources to maintain and update rules and templates. The machine learning-based model has the disadvantages of low accuracy of the hidden Markov model (HMM) and easy overfitting of the random forest algorithm. The deep learning-based model has problems such as inability to extract sequence information and gradient explosion. The main purpose of named entity recognition in the field of aviation intelligence is to identify entity information related to aviation intelligence, such as airspace names, airspace coordinates, routes, etc. There are many proper nouns, difficulties in mixed representation of entity full names and abbreviations, and lack of large-scale training data sets in the field of aviation intelligence. The CRF model can flexibly capture contextual information and correlation within the sequence in the NER task, and considers the global annotation of the sequence, which is not affected by the accumulation of local annotation errors that may be caused by manual annotation.

[0004] Faced with the large amount of aviation data compilation table information in the field of aviation intelligence, the relevant data and information are currently searched and interpreted manually. In addition, the data structure of the global aviation data compilation table is not unified, and there are bilingual data. It is very difficult to extract the structured information of the table information. Such methods often require a lot of manpower and material resources, and are prone to errors and long search time. They cannot efficiently and accurately search and interpret the relevant data, which may affect aviation operation decisions. Summary of the invention

[0005] The present invention provides a structured extraction method for aviation intelligence table information based on ontology drive, which solves the problems that the existing manual search for table information in aviation intelligence field data is time-consuming, laborious, error-prone and not automated enough.

[0006] In order to solve the above technical problems, the technical solution of the present invention is: a structured extraction method of aviation intelligence table information based on ontology drive, comprising the following steps:

[0007] S1. Construct an ontology framework model of aviation intelligence data based on the aviation information exchange model AIXM;

[0008] S2. Extract structured information from the aviation intelligence data to obtain aviation intelligence table data;

[0009] S3. Based on the ontology framework model, a regular expression matching template is designed, and the English structured text data in the aviation information form data is extracted through the regular expression in the regular expression matching template to obtain the structured entity;

[0010] S4. Construct a conditional random field model CRF to perform named entity recognition on English semi-structured and unstructured text data in aviation intelligence form data to obtain semi-structured and unstructured entities;

[0011] S5. Summarize the structured entities and semi-structured and unstructured entities to obtain the characteristic entities of aviation intelligence data and complete the structured extraction of aviation intelligence form information.

[0012] The beneficial effects of the present invention are as follows: the present invention provides an ontology framework model based on the aviation information exchange model AIXM, determines the category of entity recognition in the data through the ontology framework model, and respectively uses regular expressions and conditional random field models CRF to extract feature entities, realizes automatic information extraction of aviation information compilation tabular data to form aviation intelligence data that conforms to the AIXM structure, solves the technical problems of the existing manual search for tabular information in aviation intelligence field data that is time-consuming, labor-intensive, error-prone, and not automated enough, and effectively realizes the structured information extraction of tabular information in the aviation intelligence field.

[0013] Furthermore, the specific steps of step S2 are: using pdfplumber to parse the PDF table document of aviation intelligence data, perform document layout analysis, detect and extract the structure and text content of the table data, and obtain aviation intelligence table data.

[0014] The beneficial effects of the above further scheme are: combining document artificial intelligence technology to use pdfplumber to parse PDF table documents, analyze the document layout, detect and identify the tables in the document, reconstruct the table structure, and extract the table content. It effectively solves problems such as table spanning pages and multiple table headers, and realizes the extraction of the structure and content of table data in PDF documents.

[0015] Furthermore, the specific steps of step S4 are:

[0016] S41, pre-processing the aviation data compilation to obtain tabular data of the control area, terminal area and approach area in the aviation data compilation;

[0017] S42 uses the BIO tagging method to perform sequence tagging on the table data of the control area, terminal area, and approach area, where B represents the beginning of the entity, I represents the middle position of the entity, and O represents a non-entity;

[0018] S43. Design feature templates based on the table data structure and content characteristics of the aviation information compilation;

[0019] S44, defining a feature template, and determining a feature function of a conditional random field model CRF through the feature template;

[0020] S45, extracting and encoding different features of the words in the table data to obtain a feature vector for each word, and forming a feature matrix X through the feature vector of each word, and forming a label matrix Y through the named entity label of each word;

[0021] S46, taking the feature matrix X as the input sequence of the conditional random field model CRF, and taking the label matrix Y as the output sequence of the conditional random field model CRF;

[0022] S47, based on the input sequence and the output sequence, select the LBFGS optimization algorithm to train the conditional random field model CRF, and update the weight of the feature function to obtain the trained conditional random field model CRF;

[0023] S48, inputting the English semi-structured and unstructured text data in the aviation information form data into the trained conditional random field model CRF to obtain the conditional probability P(Y|X) corresponding to each word;

[0024] S49, based on the conditional probability P(Y|X) corresponding to each word, decode it through the Viterbi algorithm, and calculate the optimal probability of the label corresponding to each word in the text, and obtain the optimal label sequence of the text, that is, obtain the semi-structured entity

[0025] The beneficial effects of the above further scheme are: by adopting the BIO labeling method, the table data of the control area, terminal area and approach area are respectively labeled in sequence, which solves the problem of no AIP training data in the field of aviation intelligence and can effectively train the model. According to the table data structure and content characteristics of the aviation data compilation, by designing feature templates and using the CRF model to predict the text sequence labeling of the input text, the sequence labeling of entities can be effectively realized, feature entity extraction can be performed, and automatic information extraction can be realized to form aviation intelligence data that conforms to the AIXM structure.

[0026] Furthermore, the feature template in step S44 includes features and descriptions corresponding to the features;

[0027] The features include word.lower, word[-3:], word[-2:], word.isupper, word.istitle, word.isdigit, -1:word.lower, -1:word.istitle, -1:word.isupper, +1:word.lower, +1:word.istitle and +1:word.isupper, where word.lower represents the lowercase form of the word, word[-3:] represents the last three characters of the word, word[-2:] represents the last two characters of the word, word.i supper indicates whether the word is all uppercase, word.istitle indicates whether the word starts with an uppercase letter, word.isdigit indicates whether the word is a number, -1:word.lower indicates the lowercase form of the previous word, -1:word.istitle indicates whether the previous word starts with an uppercase letter, -1:word.isupper indicates whether the previous word is all uppercase, +1:word.lower indicates the lowercase form of the next word, +1:word.istitle indicates whether the next word starts with an uppercase letter and +1:word.isupper indicates whether the next word is all uppercase.

[0028] The beneficial effect of the above further scheme is: through the set feature template, the word-level features, sentence-level features and context features are extracted respectively, the entity recognition boundary is determined according to the different features, and the feature function is determined as the feature condition.

[0029] Furthermore, the expression of the characteristic function in step S44 is:

[0030]

[0031] Z(x)=v y {v i,k λ k t k (y i-1 ,y i ,x,i)+∑ i,l μ l s l (y i ,x,i)}

[0032] Among them, P(y|x) represents the conditional probability that the label matrix Y takes the value of y under the condition that the feature matrix X takes the value of x, and Z(x) represents the normalization factor. represents the exponential function, λ k represents the weight of the transfer feature function, t k (y i-1 ,y i ,x,i) represents the transition feature function defined on the edge between i and i-1, μ l represents the weight of the state characteristic function, s l (y i ,x,i) represents the state characteristic function defined on the node. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 The present invention is a flowchart of the structured extraction method of aviation intelligence table information based on ontology drive.

[0034] Figure 2 It is a schematic diagram of the structure of the ontology framework model of the present invention.

[0035] Figure 3 This is a comparison chart of the results of the two models of the present invention. DETAILED DESCRIPTION

[0036] Those skilled in the art will appreciate that the embodiments described herein are intended to help readers understand the principles of the present invention, and should be understood that the protection scope of the present invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific variations and combinations that do not deviate from the essence of the present invention based on the technical revelations disclosed by the present invention, and these variations and combinations are still within the protection scope of the present invention.

[0037] Example

[0038] like Figure 1 As shown, the present invention provides a method for extracting structured information from aviation intelligence tables based on ontology drive, comprising the following steps:

[0039] S1. Construct an ontology framework model of aviation intelligence data based on the aviation information exchange model AIXM;

[0040] S2. Extract structured information from the aviation intelligence data to obtain aviation intelligence table data;

[0041] S3. Based on the ontology framework model, a regular expression matching template is designed, and the English structured text data in the aviation information form data is extracted through the regular expression in the regular expression matching template to obtain the structured entity;

[0042] S4. Construct a conditional random field model CRF to perform named entity recognition on English semi-structured and unstructured text data in aviation intelligence form data to obtain semi-structured and unstructured entities;

[0043] S5. Summarize the structured entities and semi-structured and unstructured entities to obtain the characteristic entities of aviation intelligence data and complete the structured extraction of aviation intelligence form information.

[0044] In this embodiment, before extracting structured information, it is necessary to determine which categories of entities to extract, so an ontology framework model needs to be constructed. This embodiment constructs a resource description framework RDF, ie, an ontology framework model, based on the aviation information exchange model AIXM of a global standardized structure.

[0045] like Figure 2 As shown in the figure, the ontology framework model structure diagram takes airspace aviation intelligence data as an example. By introducing the ontology framework model and combining it with the AIXM model, the knowledge in the aviation intelligence field is standardizedly described.

[0046] The specific steps of step S2 are: using pdfplumber to parse the PDF table document of aviation intelligence data, perform document layout analysis, detect and extract the structure and text content of the table data, and obtain aviation intelligence table data.

[0047] In this embodiment, pdfplumber is first used to parse the PDF table document of aviation intelligence data, that is, the aviation intelligence data parsed into a PDF file, and the document layout is analyzed to detect and extract the table data structure and its text content. Regular expressions are used to extract the English structured text data that conforms to the AIXM standard from the processed table data to obtain structured entities.

[0048] Taking the control area and route table data as an example, a regular expression matching template is designed based on the ontology framework model, as shown in Table 1. Table 1 is an example table of some regular expression matching templates.

[0049] Table 1

[0050]

[0051] In this embodiment, English structured text data that conforms to the AIXM standard can use regular expressions to extract entities, while English semi-structured and unstructured text data that do not conform to the AIXM standard cannot use regular expressions to extract entities. Therefore, a conditional random field model CRF is designed to perform named entity recognition and extract feature entities.

[0052] The specific steps of step S4 are:

[0053] S41, pre-processing the aviation data compilation to obtain tabular data of the control area, terminal area and approach area in the aviation data compilation;

[0054] S42 uses the BIO tagging method to perform sequence tagging on the table data of the control area, terminal area, and approach area, where B represents the beginning of the entity, I represents the middle position of the entity, and O represents a non-entity;

[0055] S43. Design feature templates based on the table data structure and content characteristics of the aviation information compilation;

[0056] S44, defining a feature template, and determining a feature function of a conditional random field model CRF through the feature template;

[0057] S45, extracting and encoding different features of the words in the table data to obtain a feature vector for each word, and forming a feature matrix X through the feature vector of each word, and forming a label matrix Y through the named entity label of each word;

[0058] S46, taking the feature matrix X as the input sequence of the conditional random field model CRF, and taking the label matrix Y as the output sequence of the conditional random field model CRF;

[0059] S47, based on the input sequence and the output sequence, select the LBFGS optimization algorithm to train the conditional random field model CRF, and update the weight of the feature function to obtain the trained conditional random field model CRF;

[0060] S48, inputting the English semi-structured and unstructured text data in the aviation information form data into the trained conditional random field model CRF to obtain the conditional probability P(Y|X) corresponding to each word;

[0061] S49, based on the conditional probability P(Y|X) corresponding to each word, decode it through the Viterbi algorithm, and calculate the optimal probability of the label corresponding to each word in the text, and obtain the optimal label sequence of the text, that is, obtain the semi-structured entity

[0062] The feature template in step S44 includes features and descriptions corresponding to the features;

[0063] The features include word.lower, word[-3:], word[-2:], word.isupper, word.istitle, word.isdigit, -1:word.lower, -1:word.istitle, -1:word.isupper, +1:word.lower, +1:word.istitle and +1:word.isupper, where word.lower represents the lowercase form of the word, word[-3:] represents the last three characters of the word, word[-2:] represents the last two characters of the word, word.i supper indicates whether the word is all uppercase, word.istitle indicates whether the word starts with an uppercase letter, word.isdigit indicates whether the word is a number, -1:word.lower indicates the lowercase form of the previous word, -1:word.istitle indicates whether the previous word starts with an uppercase letter, -1:word.isupper indicates whether the previous word is all uppercase, +1:word.lower indicates the lowercase form of the next word, +1:word.istitle indicates whether the next word starts with an uppercase letter and +1:word.isupper indicates whether the next word is all uppercase.

[0064] The expression of the characteristic function in step S44 is:

[0065]

[0066] Z(x)=∑ y {∑ i,k λ k t k (y i-1 ,y i ,x,i)+∑ i,l μ l s l (y i ,x,i)}

[0067] Among them, P(y|x) represents the conditional probability that the label matrix Y takes the value of y under the condition that the feature matrix X takes the value of x, Z(x) represents the normalization factor, exp() represents the exponential function, and λ k represents the weight of the transfer feature function, t k (y i-1 ,y i ,x,i) represents the transition feature function defined on the edge between i and i-1, μ l represents the weight of the state characteristic function, s l (y i,x,i) represents the state characteristic function defined on the node.

[0068] In this embodiment, the transition feature function and the state feature function are defined to have values ​​of 1 or 0; when the feature condition is met, the value is 1, otherwise it is 0. In the CRF model, the selection of feature templates can determine the feature function, and a suitable feature template can provide a unified pattern for the generation of feature functions.

[0069] In view of the data structure and content characteristics of the aviation data compilation table, this embodiment selects and designs a feature template to determine the feature function, as shown in Table 2, which is a feature template.

[0070] Table 2

[0071]

[0072] In this embodiment, the tabular data of the control area, terminal area and approach area in the aviation data compilation after sequence annotation is used as a training set. Each sample is a text sequence, and each sequence has a corresponding label for indicating the starting position and entity type of the entity.

[0073] Taking the text data in the table as the input sequence, this embodiment takes "12 500m—" in "ZBAAAR01 N432658E1112149N413057E1125726N404413E1125425N414843E1140800N415205E1145015N422553E1151044N423342E1175429N431442E1173100N452317E1152115National Boundaries N432658E111214912 500m——————7 800m(exclusive)" as an example, and calculates the conditional probability P(y) of its corresponding label "B-upperlimit I-upperlimit I-upperlimit". 24 ,y 25 ,y 26 |x), and obtain the optimal path through Viterbi algorithm decoding to obtain the optimal label sequence.

[0074] This embodiment is described by comparing the random forest algorithm with the method of the present invention. When the random forest is used to perform named entity recognition on the aviation data compilation, the feature array is converted based on the feature template in Table 2 before the named entity recognition is performed. The random forest model is trained by the tabular data of the control area, terminal area and approach area in the aviation data compilation, and a sub-data set is randomly extracted from the original data set to construct a single decision tree. Each decision tree splits the node by randomly selecting the feature space, and finally the prediction result is obtained by the voting results of all single decision trees.

[0075] In this embodiment, the experimental parameters are set as follows: for the conditional random field model, the LBFGS algorithm is used as the optimization algorithm, the coefficients of the L1 and L2 regularization terms are set to 0.1, and the maximum number of iterations is limited to 200 times for training and prediction of the two data sets. For the random forest algorithm, the parameters are set to 20 decision trees. Both models evaluate the data sets using 5-fold cross validation. The F1 value is used as the model effect evaluation index. The commonly used evaluation indicators in the NER task are accuracy P, recall R and F1 value, and the calculation formula is:

[0076]

[0077]

[0078]

[0079] Among them, T P represents the number of correct matches, F N Indicates the number of positive classes predicted as negative classes, F P Indicates the number of negative classes predicted as positive classes.

[0080] The experiment uses precision P, recall R and F1 value as evaluation indicators, and compares the experimental effects of the two models under the same data sets (terminal area and approach area data sets, control area data sets). The results are shown in Figure 3. Table 3 is a comparison chart of the results.

[0081]

[0082] like Figure 3 As shown, a comparison diagram of precision P, recall R and F1 value is drawn. From the above, it can be seen that on the control area data set: when the conditional random field model CRF and the random forest model are used to extract the features of the control area, the extraction effect based on the conditional random field model CRF is better than the random forest algorithm, and its F1 value reaches 87%, while the F1 value extracted based on the random forest algorithm is 73%.

[0083] On the terminal area and approach area data sets, when the conditional random field model CRF and the random forest model are used to extract features of the terminal area and approach area, the extraction effect based on the conditional random field model CRF is also better than the random forest algorithm, with its F1 value reaching 77%, while the F1 value based on the random forest algorithm is 50%.

[0084] In summary, the overall extraction effect of the method of the present invention on the data set (terminal area and approach area data set, control area data set) is better than that of the random forest algorithm, and it has a higher extraction effect. The named entity recognition effect of entities of various label types is relatively ideal, and it is suitable for the extraction of aviation intelligence form data.

[0085] Therefore, the present invention constructs an ontology framework model based on AIXM, adopts regular expressions and conditional random fields (CRF) to extract feature entities, realizes automatic information extraction of aviation information compilation tabular data to form aviation intelligence data that conforms to the AIXM structure, solves the technical problems of the existing manual search for tabular information in aviation intelligence data, such as time-consuming, laborious, and error-prone, and is not automated enough, and effectively realizes the structured information extraction of tabular information in the aviation intelligence field.

Claims

1. A structured extraction method of aviation intelligence table information based on ontology drive, characterized in that: The following steps are involved: S1. Construct an ontology framework model of aviation intelligence data based on the aviation information exchange model AIXM; S2. Extract structured information from the aviation intelligence data to obtain aviation intelligence table data; S3. Based on the ontology framework model, a regular expression matching template is designed, and the English structured text data in the aviation information form data is extracted through the regular expression in the regular expression matching template to obtain the structured entity; S4. Construct a conditional random field model CRF to perform named entity recognition on English semi-structured and unstructured text data in aviation intelligence form data to obtain semi-structured and unstructured entities; S5. Summarize the structured entities and semi-structured and unstructured entities to obtain the characteristic entities of aviation intelligence data and complete the structured extraction of aviation intelligence form information.

2. The ontology-driven structured extraction method of aviation intelligence table information according to claim 1 is characterized in that: The specific steps of step S2 are: using pdfplumber to parse the PDF table document of aviation intelligence data, perform document layout analysis, detect and extract the structure and text content of the table data, and obtain aviation intelligence table data.

3. The ontology-driven structured extraction method of aviation intelligence table information according to claim 1 is characterized in that: The specific steps of step S4 are: S41, pre-processing the aviation data compilation to obtain tabular data of the control area, terminal area and approach area in the aviation data compilation; S42 uses the BIO tagging method to perform sequence tagging on the table data of the control area, terminal area, and approach area, where B represents the beginning of the entity, I represents the middle position of the entity, and O represents a non-entity; S43. Design feature templates based on the table data structure and content characteristics of the aviation information compilation; S44, defining a feature template, and determining a feature function of a conditional random field model CRF through the feature template; S45, extracting and encoding different features of the words in the table data to obtain a feature vector for each word, and forming a feature matrix X through the feature vector of each word, and forming a label matrix Y through the named entity label of each word; S46, taking the feature matrix X as the input sequence of the conditional random field model CRF, and taking the label matrix Y as the output sequence of the conditional random field model CRF; S47, based on the input sequence and the output sequence, select the LBFGS optimization algorithm to train the conditional random field model CRF, and update the weight of the feature function to obtain the trained conditional random field model CRF; S48, inputting the English semi-structured and unstructured text data in the aviation information form data into the trained conditional random field model CRF to obtain the conditional probability P(Y|X) corresponding to each word; S49. Based on the conditional probability P(Y|X) corresponding to each word, decoding is performed through the Viterbi algorithm, and the optimal probability of the label corresponding to each word in the text is calculated to obtain the optimal label sequence of the text, that is, to obtain a semi-structured entity.

4. The ontology-driven structured extraction method of aviation intelligence table information according to claim 3 is characterized in that: The feature template in step S44 includes features and descriptions corresponding to the features; The features include word.lower, word[-3:], word[-2:], word.isupper, word.istitle, word.isdigit, -1:word.lower, -1:word.istitle, -1:word.isupper, +1:word.lower, +1:word.istitle and +1:word.isupper, where word.lower represents the lowercase form of the word, word[-3:] represents the last three characters of the word, word[-2:] represents the last two characters of the word, word.i supper indicates whether the word is all uppercase, word.istitle indicates whether the word starts with an uppercase letter, word.isdigit indicates whether the word is a number, -1:word.lower indicates the lowercase form of the previous word, -1:word.istitle indicates whether the previous word starts with an uppercase letter, -1:word.isupper indicates whether the previous word is all uppercase, +1:word.lower indicates the lowercase form of the next word, +1:word.istitle indicates whether the next word starts with an uppercase letter and +1:word.isupper indicates whether the next word is all uppercase.

5. The ontology-driven structured extraction method of aviation intelligence table information according to claim 3 is characterized in that: The expression of the characteristic function in step S44 is: Z(x)=∑ y {∑ i,k λ k t k (y i-1 ,y i ,x,i)+∑ i,l μ l s l (y i ,x,i)} Among them, P(y|x) represents the conditional probability that the label matrix Y takes the value of y under the condition that the feature matrix X takes the value of x, and Z(x) represents the normalization factor. represents the exponential function, λ k represents the weight of the transfer feature function, t k (y i-1 ,y i ,x,i) represents the transition feature function defined on the edge between i and i-1, μ l represents the weight of the state characteristic function, s l (y i ,x,i) represents the state characteristic function defined on the node.

Images