Entity recognition-oriented aeronautical intelligence field corpus construction method

Abstract

The invention provides an entity recognition-oriented aviation intelligence field corpus construction method, which relates to the field of data processing, and comprises the following steps: acquiring aviation intelligence data, and normalizing the original corpus of aviation intelligence; in combination with an aviation intelligence exchange model AIXM, constructing a domain corpus labeling system and forming a labeling criterion, and labeling to obtain an aviation intelligence labeling corpus; and in the construction process, superposition of a sequence labeling layer is carried out based on a regular template, vocabulary feature information is incorporated into character representation, a character representation layer is adjusted to introduce dictionary information, a bidirectional long-short term memory network model is adopted to carry out entity recognition, and construction of the corpus in the aviation intelligence field is completed. The technical problems that no industrial special corpus exists in the current aviation field, and data updating and processing depend on manual work are solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of data processing, and in particular relates to a method for constructing a corpus in the field of aviation intelligence for entity recognition. Background Art

[0002] Aviation intelligence services are a crucial support service for aviation operations, providing aviation data and operational decision-making information for civil aviation transport, encompassing a vast amount of relevant domain knowledge and specialized vocabulary. With the rapid development of global aviation operations, a vast amount of information impacting aviation planning and daily operations urgently needs to be digitized. The foundation for efficient digital transformation lies in the construction of a domain corpus.

[0003] Traditional aviation intelligence work relies primarily on manual interpretation, lacking structured processing and value mining of large amounts of intelligence data. With the rapid development of global aviation operations, the updating of aviation information has resulted in a massive amount of information that impacts aviation planning and daily operations, hindering the development and application of further information platforms such as big data mining, pattern analysis, and decision support. To address this issue, manual methods are often used to collect and update information, requiring significant manpower and resources. Manual interpretation is inefficient and prone to errors, hindering the structuring and digitization of aviation intelligence products. Summary of the Invention

[0004] In response to the above-mentioned deficiencies in the existing technology, the present invention provides a method for constructing an aviation intelligence corpus for entity recognition, which solves the technical problem that there is currently no industry-specific corpus in the aviation field and that data updating and processing rely on manual labor.

[0005] In order to achieve the above objectives, the technical solution adopted by the present invention is: a method for constructing an aviation intelligence corpus for entity recognition, characterized by comprising the following steps:

[0006] S1. Obtain aviation intelligence data and perform standardization on the original aviation intelligence data;

[0007] S2. Combined with the Aviation Information Exchange Model (AIXM), a domain corpus annotation system is constructed and annotation criteria are formed. The aviation information annotated corpus is annotated to obtain the aviation information annotated corpus. The Aviation Information Exchange Model (AIXM) displays the attributes of the entity parts and the relationships between the entities.

[0008] S3. Based on the annotated aviation intelligence annotated corpus, construct an entity dictionary of abbreviations in the aviation intelligence field. During the construction process, a sequence annotation layer is superimposed based on a regular template, lexical feature information is incorporated into the character representation, and the character representation layer is adjusted to introduce dictionary information.

[0009] S4. Based on the abbreviation entity dictionary in the aviation intelligence field, a bidirectional long short-term memory network model is used for entity recognition to complete the construction of the aviation intelligence field corpus.

[0010] The beneficial effects of the present invention are as follows: the present invention adopts the BiLSTM-CRF model (bidirectional long short-term memory network) to integrate the aviation intelligence product abbreviation dictionary, which solves the technical problem that there is no industry-specific corpus in the current aviation field and relies on manual data updating and processing. It effectively improves the recognition effect of named entities in the aviation intelligence corpus and provides support for subsequent entity recognition of aviation intelligence information and large-scale corpus construction.

[0011] Furthermore, the S1 is specifically:

[0012] Obtain aviation intelligence information;

[0013] Based on aviation intelligence data, Chinese and English parallel corpora of the original aviation intelligence corpus are extracted, and the extracted Chinese and English parallel corpora are subjected to text parsing and normalization processing.

[0014] The beneficial effect of the above further scheme is that corpora aligned at different levels are of great significance to the construction of domain language resources and are one of the main directions in natural language processing research. Exploring the alignment of Chinese and English sentences in aviation intelligence data is of practical significance for natural language processing tasks such as the migration application of domain corpora, the realization of parallel corpus construction, and intelligent machine translation.

[0015] Furthermore, the S2 is specifically:

[0016] Based on the standardized description of domain knowledge in the Aeronautical Information Exchange Model (AIXM), a domain corpus annotation system and annotation guidelines are constructed. The AIXM model displays the attributes of entities and the relationships between entities. Based on the AIXM standard, a domain corpus annotation system is developed based on entity recognition requirements to cover the entity categories in the aeronautical information data compilation.

[0017] The BIO sequence annotation method is used, and the doccano online annotation platform is used to perform entity sequence annotation. After the entity sequence is annotated, the sequence annotation results are output in jsonl format and converted into csv sequence annotation data;

[0018] Based on the sequence labeling results, each character in the aviation intelligence text is mapped to the corresponding entity category and sequence position to obtain the aviation intelligence annotated corpus, where B represents the beginning of the entity, I represents the inside of the entity, and O represents the component of the non-entity vocabulary.

[0019] The beneficial effect of this further approach is that the core task of corpus construction is to develop a comprehensive annotation system and annotation guidelines. In general corpora, entity categories typically include names of people, places, and organizations, which are widely applicable to various named entity recognition tasks. However, in the civil aviation field, general entity labels are difficult to cover the large number of specialized terms. Therefore, it is necessary to establish a annotation system and guidelines based on semantically standardized representation to serve aviation intelligence corpora.

[0020] Furthermore, the S3 is specifically:

[0021] A method combining a thesaurus and rule statistics was used to extract and filter PDF documents from the aviation information product abbreviation list in the thesaurus. The obtained abbreviation pairs were divided according to feature types, and regular expressions were used to establish combination rules for template matching. Sequence annotation layers were superimposed based on regular templates, lexical feature information was incorporated into character representation, and the character representation layer was adjusted to introduce dictionary information.

[0022] Based on the division results, the mapping between abbreviations and original entities is achieved by constructing abbreviation pairs;

[0023] Based on the mapping results, the original entity combination rules and the annotation sequences in the aviation intelligence annotation corpus are integrated into the aviation intelligence tag corpus as lexical features to complete the construction of the abbreviation entity dictionary in the aviation intelligence field.

[0024] The beneficial effect of the above further scheme is: due to the particularity of civil aviation professional terms, professional technical terms or civil aviation abbreviations are often used in various information expressions to express them, which makes it difficult to identify domain terms. By capturing the vocabulary features through the semantic understanding of contextual vocabulary and realizing the feature fusion of domain vocabulary, the model can significantly improve the entity recognition accuracy of aviation professional terms.

[0025] Furthermore, the S4 is specifically:

[0026] Take a piece of aviation information from the aviation information annotated corpus that incorporates lexical features as input and convert it into a word vector using the input layer;

[0027] The word vector is input into the bidirectional long short-term memory network BiLSTM, and the output sequence is obtained by combining the hidden states of the forward layer and the backward layer at each position;

[0028] The output sequence is decoded using the conditional random field layer CRF, mapped to K dimensions, and normalized using the softmax function;

[0029] Based on the label corresponding to the maximum value of the normalized calculation result, the optimal output sequence label is obtained, and the entity recognition result is obtained, completing the construction of the aviation intelligence field corpus.

[0030] The beneficial effects of the above further scheme are: the BiLSTM layer can automatically learn the contextual features of the corpus, introduce the CRF layer to learn the constraints of sequence labeling, consider the sequence between output labels by transferring features to ensure the validity of the prediction results, and use an external dictionary to improve the performance of the domain entity recognition task through the dictionary without changing the model architecture.

[0031] Furthermore, the expression of the output sequence is as follows:

[0032] p i =g(Wh i +Wh i ')

[0033] h i '=f(U'X i +V'h' t-1 )

[0034] h i =f(UX i +Vh t-1 )

[0035] Among them, p i Represents the output sequence, g represents the activation function, which acts on the hidden layer to the output layer, h i and h i 'represent the output of the forward layer and the backward layer respectively, f represents the activation function, which acts on the hidden layer, W, U, V, U' and V' all represent weights, X i Represents the i-th character of the input sequence, h i-1 represents the output of the forward LSTM at time t-1, h' i-1 Represents the output of the backward layer at time t-1.

[0036] The beneficial effect of the above further scheme is that by splicing the forward and backward hidden states, it is possible to combine contextual information, capture the forward and backward dependencies of the text in both directions, and improve the model's understanding and recognition ability of complex sentences and professional terms.

[0037] Furthermore, the expression of the entity recognition result is as follows:

[0038] Y * =argmax S(X,Y)

[0039]

[0040] Among them, i represents the sequence position index, n represents the input sequence length, Represents the label y i to y i+1 The transfer score, represents the input x at position i i Marked as y i The score of the label sequence is the emission score, S(X,Y) represents the combined score of the emission score and the transfer score, Y * represents the optimal label sequence of the output.

[0041] The beneficial effect of the above further scheme is that the scoring function outputs the global optimal label sequence by jointly optimizing the emission score and the transfer score, thereby strengthening the global optimization of the label sequence, ensuring the legitimacy of the label sequence, and improving the consistency between labels. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 Flow chart of the method of the present invention.

[0043] Figure 2 This is an example diagram of the UML structure of airport-class entities and attributes in the Aviation Information Exchange Model (AIXM) referred to in the present invention.

[0044] Figure 3 This is a schematic diagram of the Bi-LSTM-CRF model structure used in the present invention. DETAILED DESCRIPTION

[0045] The specific embodiments of the present invention are described below to facilitate understanding of the present invention by those skilled in the art. However, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the appended claims, these changes are obvious, and all inventions and creations utilizing the concepts of the present invention are protected.

[0046] Example

[0047] The basic idea of ​​this paper is to propose a method for constructing an aviation information domain corpus for entity recognition based on the characteristics of text data structure in aviation data compilation, combining the Aeronautical Information Exchange Model (AIXM) and named entity recognition technology. Based on the definition and classification of aviation elements in the Aeronautical Information Exchange Model (AIXM), an entity annotation system for aviation information domain is established, annotation standards are formulated, and characteristic entities are extracted from aviation information product information under this standard. Figure 1 As shown, the present invention provides a method for constructing a corpus in the field of aviation intelligence for entity recognition, and its implementation method is as follows:

[0048] S1. Obtain aviation intelligence data and normalize the original aviation intelligence data:

[0049] Obtain aviation intelligence information;

[0050] Based on aviation intelligence data, Chinese and English parallel corpora of the original aviation intelligence corpus are extracted, and the extracted Chinese and English parallel corpora are subjected to text parsing and normalization processing.

[0051] In this embodiment, for the sampled aviation intelligence data, a text extraction and normalization algorithm based on coordinate block positioning is used to extract Chinese and English parallel corpora, and the corpora are subjected to text parsing and normalization processing to ensure the quality of corpus construction and entity recognition.

[0052] S2. Based on the Aviation Information Exchange Model (AIXM), a domain corpus annotation system is constructed and annotation criteria are formed. The aviation information annotated corpus is annotated to obtain the aviation information annotated corpus. The Aviation Information Exchange Model (AIXM) displays the attributes of entities and the relationships between entities. The implementation method is as follows:

[0053] Based on the standardized description of domain knowledge in the Aeronautical Information Exchange Model (AIXM), a domain corpus annotation system and annotation guidelines are constructed. The AIXM model displays the attributes of entities and the relationships between entities. Based on the AIXM standard, a domain corpus annotation system is developed based on entity recognition requirements to cover the entity categories in the aeronautical information data compilation.

[0054] The BIO sequence annotation method is used, and the doccano online annotation platform is used to perform entity sequence annotation. After the entity sequence is annotated, the sequence annotation results are output in jsonl format and converted into csv sequence annotation data;

[0055] Based on the sequence labeling results, each character in the aviation intelligence text is mapped to the corresponding entity category and sequence position to obtain the aviation intelligence annotated corpus, where B represents the beginning of the entity, I represents the inside of the entity, and O represents the component of the non-entity vocabulary.

[0056] In this embodiment, the Aeronautical Information Exchange Model (AIXM) consists of the Aeronautical Information Conceptual Model (AICM) and the AIXM. The AICM standardizes the attribute and relationship definitions of aeronautical information, categorizing aeronautical information into six parts: airports, airspace, points, routes, procedures, and services. It also provides clear hierarchical classification and definitions for each element, making it an internationally recognized representation of aviation entities.

[0057] In this embodiment, the {B, I, O} sequence annotation, namely the "beginning-inside-non-entity" annotation, is a annotation method used for tasks such as named entity recognition in natural language processing, where B represents the beginning of the entity, I represents the inside of the entity, and O represents a component of the non-entity vocabulary. Its core is to determine the sequence position of each word in the text to be annotated.

[0058] In this embodiment, a tagging system is constructed based on the Aviation Information Exchange Model (AIXM) specification and a tagging criterion is formed, and a large amount of annotated corpus is obtained by using BIO sequence tagging. Figure 2 The UML diagram shown shows the attributes of entities such as airports, runways, and organizations in the Aeronautical Information Exchange Model (AIXM) and their relationships.

[0059] Under the Aeronautical Information Exchange Model (AIXM) standard, the following annotation system is developed based on the need for entity identification, covering the entity categories in the aeronautical information data compilation:

[0060] Table 1

[0061]

[0062] S3. Based on the annotated aviation intelligence corpus, a dictionary of abbreviations in the aviation intelligence field is constructed. During the construction process, a sequence annotation layer is superimposed based on a regular template, lexical feature information is incorporated into the character representation, and the character representation layer is adjusted to introduce dictionary information without changing the original architecture of the model. The implementation method is as follows:

[0063] A method combining a thesaurus and rule statistics was used to extract and filter PDF documents from the aviation information product abbreviation list in the thesaurus. The obtained abbreviation pairs were divided according to feature types, and regular expressions were used to establish combination rules for template matching. Sequence annotation layers were superimposed based on regular templates, lexical feature information was incorporated into character representation, and the character representation layer was adjusted to introduce dictionary information.

[0064] Based on the division results, the mapping between abbreviations and original entities is achieved by constructing abbreviation pairs;

[0065] Based on the mapping results, the original entity combination rules and the annotation sequences in the aviation intelligence annotation corpus are integrated into the aviation intelligence tag corpus as lexical features to complete the construction of the abbreviation entity dictionary in the aviation intelligence field.

[0066] In this embodiment, a method based on a combination of a vocabulary and rule statistics is adopted. The abbreviation table of aviation information products in the General Rules for Aeronautical Information Compilation 2.2 (GEN 2.2) is used to extract and filter PDF documents using a method based on coordinate block positioning. The obtained abbreviation pairs are divided according to feature types, and regular expressions are used to establish combination rules for template matching. Based on the division results, the mapping between abbreviations and original entities is achieved by constructing abbreviation pairs.

[0067] The coordinate block positioning method is one of the current mainstream paragraph alignment methods. The specific steps of the aviation intelligence corpus extraction method based on coordinate block positioning are as follows:

[0068] Step 1: Traverse each page in the original PDF corpus and obtain the page rectangular area;

[0069] Step 2: Position the coordinate block according to the top and bottom information positions and the center line position of the page;

[0070] Step 3: Crop to remove top and bottom interference information, and output the texts on the left and right as Chinese and English corpora respectively;

[0071] Step 4: Use regular expressions to obtain the required text corpus resources after removing table information.

[0072] In this example, a dictionary of abbreviations in the aviation intelligence field was constructed. Using a method combining a vocabulary and rule statistics, the abbreviations used in aviation intelligence products (GEN2.2) were extracted and filtered. The resulting 1,145 abbreviation pairs were divided into three categories based on their feature types, as shown in Table 2:

[0073] Table 2

[0074]

[0075]

[0076] By establishing abbreviation pairs or internal rules to complete the mapping between abbreviations and original entities, and integrating their combination rules and their annotation sequences as lexical features into the corpus, the extraction task of entities containing abbreviations is realized, effectively improving the entity recognition effect of the model.

[0077] S4. Based on the abbreviation entity dictionary in the aviation intelligence field, a bidirectional long short-term memory network model is used for entity recognition to complete the construction of the aviation intelligence field corpus. The implementation method is as follows:

[0078] Take a piece of aviation information from the aviation information annotated corpus that incorporates lexical features as input and convert it into a word vector using the input layer;

[0079] The word vector is input into the bidirectional long short-term memory network BiLSTM, and the output sequence is obtained by combining the hidden states of the forward layer and the backward layer at each position;

[0080] The output sequence is decoded using the conditional random field layer CRF, mapped to K dimensions, and normalized using the softmax function;

[0081] Based on the label corresponding to the maximum value of the normalized calculation result, the optimal output sequence label is obtained, and the entity recognition result is obtained, completing the construction of the aviation intelligence field corpus.

[0082] In this embodiment, the aviation intelligence annotated corpus obtained in step S2 is integrated with vocabulary features and an entity recognition experiment is performed using a BiLSTM-CRF model to test the quality of the self-built corpus. The principle of BiLSTM-CRF entity recognition is as follows:

[0083] like Figure 3 As shown, an aviation information corpus is used as the input sequence X=(X1,X2,…,X j ) as an example, X j Represents the jth input sequence, which is converted into a word vector after the input layer. In the vector matrix n*d, n is the length of the input sequence, and d is the dimension of the Embedding layer in the model parameters. The bidirectional long short-term memory network used in the hidden layer includes two LSTM networks, one forward and one backward, which transforms the input layer vector sequence (X1, X2, ..., X j ) is input into BiLSTM, and the relevant formula is expressed as:

[0084] h i =f(UX i +Vh i-1 )

[0085] h i '=f(U'X i +V'h' t-1 )

[0086] h i =f(UX i +Vh t-1 )

[0087] Among them, p i Represents the output sequence, g represents the activation function, which acts on the hidden layer to the output layer, h i and h i 'represent the output of the forward layer and the backward layer respectively, f represents the activation function, which acts on the hidden layer, W, U, V, U' and V' all represent weights, X i Represents the i-th character of the input sequence, hi-1 represents the output of the forward LSTM at time t-1, h' i-1 Represents the output of the backward layer at time t-1, and outputs the sequence p i As the input of the CRF layer, the label sequence is decoded and the output sequence p i The dimension is mapped to the K dimension (representing the number of label categories, K = 6 in this task) and normalized by the softmax function. The calculation formula is:

[0088] Y i =softmax(ωH i +b)

[0089] Among them, ω represents the weight, b represents the bias term, and Y i Represents the probability distribution of the model's prediction for the i-th input unit label.

[0090] The final output result is Y in the above formula i The label corresponding to the maximum value of , in order to ensure the validity of the entire output sequence result, the conditional random field layer CRF is introduced to establish the rule constraints of the label. For example, the label of ORG-B must be followed by ORG-I instead of PER-I. The corresponding prediction score of the output sequence is shown in the following formula. The obtained path score S(X,Y) is the optimal output sequence label. The present invention adopts the Viterbi algorithm for decoding:

[0091]

[0092] Y * =argmax S(X,Y)

[0093] Among them, i represents the sequence position index, n represents the input sequence length, A yi+yi+1 Represents the label y i to y i+1 The transfer score, P i,yi represents the input x at position i i Marked as y i The score of the label sequence is the emission score, S(X,Y) represents the combined score of the emission score and the transfer score, Y * represents the optimal label sequence of the output.

[0094] In summary, this invention, for the first time, builds a domain-specific corpus for entity recognition tasks in the aviation intelligence field, integrating intelligence product abbreviations. This automated recognition approach improves the efficiency of manual processing of aviation intelligence data. This invention effectively utilizes massive amounts of aviation data and combines it with text features for data mining, resulting in higher extraction efficiency. By integrating the lexical features of aviation intelligence abbreviations, this invention significantly improves the model's named entity recognition performance compared to traditional NER models.

Claims

1. A method for constructing a corpus in the field of aviation intelligence for entity recognition, characterized in that: The following steps are involved: S1. Obtain aviation intelligence data and perform standardization on the original aviation intelligence data; S2. Combined with the Aviation Information Exchange Model (AIXM), a domain corpus annotation system is constructed and annotation criteria are formed. The aviation information annotated corpus is annotated to obtain the aviation information annotated corpus. The Aviation Information Exchange Model (AIXM) displays the attributes of the entity parts and the relationships between the entities. S3. Based on the annotated aviation intelligence annotated corpus, construct an entity dictionary of abbreviations in the aviation intelligence field. During the construction process, a sequence annotation layer is superimposed based on a regular template, lexical feature information is incorporated into the character representation, and the character representation layer is adjusted to introduce dictionary information. S4. Based on the abbreviation entity dictionary in the aviation intelligence field, a bidirectional long short-term memory network model is used for entity recognition to complete the construction of the aviation intelligence field corpus.

2. The method for constructing an aviation intelligence domain corpus for entity recognition according to claim 1 is characterized in that: The S1 is specifically: Obtain aviation intelligence information; Based on aviation intelligence data, Chinese and English parallel corpora of the original aviation intelligence corpus are extracted, and the extracted Chinese and English parallel corpora are subjected to text parsing and normalization processing.

3. The method for constructing an aviation intelligence domain corpus for entity recognition according to claim 1, characterized in that: The S2 is specifically: Based on the standardized description of domain knowledge in the Aeronautical Information Exchange Model (AIXM), a domain corpus annotation system and annotation guidelines are constructed. The AIXM model displays the attributes of entities and the relationships between entities. Based on the AIXM standard, a domain corpus annotation system is developed based on entity recognition requirements to cover the entity categories in the aeronautical information data compilation. The BIO sequence annotation method is used, and the doccano online annotation platform is used to perform entity sequence annotation. After the entity sequence is annotated, the sequence annotation results are output in jsonl format and converted into csv sequence annotation data; Based on the sequence labeling results, each character in the aviation intelligence text is mapped to the corresponding entity category and sequence position to obtain the aviation intelligence annotated corpus, where B represents the beginning of the entity, I represents the inside of the entity, and O represents the component of the non-entity vocabulary.

4. The method for constructing an aviation intelligence domain corpus for entity recognition according to claim 1, characterized in that: The S3 is specifically: A method combining a thesaurus and rule statistics was used to extract and filter PDF documents from the aviation information product abbreviation list in the thesaurus. The obtained abbreviation pairs were divided according to feature types, and regular expressions were used to establish combination rules for template matching. Sequence annotation layers were superimposed based on regular templates, lexical feature information was incorporated into character representation, and the character representation layer was adjusted to introduce dictionary information. Based on the division results, the mapping between abbreviations and original entities is achieved by constructing abbreviation pairs; Based on the mapping results, the original entity combination rules and the annotation sequences in the aviation intelligence annotation corpus are integrated into the aviation intelligence tag corpus as lexical features to complete the construction of the abbreviation entity dictionary in the aviation intelligence field.

5. The method for constructing an aviation intelligence domain corpus for entity recognition according to claim 4 is characterized in that: The S4 is specifically: Take a piece of aviation information from the aviation information annotated corpus that incorporates lexical features as input and convert it into a word vector using the input layer; The word vector is input into the bidirectional long short-term memory network BiLSTM, and the output sequence is obtained by combining the hidden states of the forward layer and the backward layer at each position; The output sequence is decoded using the conditional random field layer CRF, mapped to K dimensions, and normalized using the softmax function; Based on the label corresponding to the maximum value of the normalized calculation result, the optimal output sequence label is obtained, and the entity recognition result is obtained, completing the construction of the aviation intelligence field corpus.

6. The method for constructing an aviation intelligence domain corpus for entity recognition according to claim 5 is characterized in that: The expression of the output sequence is as follows: p i =g(Wh i +Wh i ') h i '=f(U'X i +V'h' t-1 ) h i =f(UX i +Vh t-1 ) Among them, p i Represents the output sequence, g represents the activation function, which acts on the hidden layer to the output layer, h i and h i 'represent the output of the forward layer and the backward layer respectively, f represents the activation function, which acts on the hidden layer, W, U, V, U' and V' all represent weights, X i Represents the i-th character of the input sequence, h i-1 represents the output of the forward LSTM at time t-1, h' i-1 Represents the output of the backward layer at time t-1.

7. The method for constructing an aviation intelligence domain corpus for entity recognition according to claim 5, characterized in that: The expression of the entity recognition result is as follows: Y * =argmax S(X,Y) Among them, i represents the sequence position index, n represents the input sequence length, Represents the label y i to y i+1 The transfer score, P i,yi represents the input x at position i i Marked as y i The score of the label sequence is the emission score, S(X,Y) represents the combined score of the emission score and the transfer score, Y * represents the optimal label sequence of the output.

Images