Road disease knowledge graph construction and entity alignment method

Abstract

The invention discloses a road disease knowledge graph construction and entity alignment method. The method comprises the following steps: S1, selecting data to construct an ontology model; s2, preprocessing the data, and extracting disease-related entities by adopting an entity recognition model; s3, judging whether the identified entity pair has a semantic relationship or not according to a relationship extraction model; s4, aligning the disease terms and the entities; and S5, storing and applying the knowledge graph. The method has the advantages that the disease ontology model is constructed, then the entity recognition model is adopted to extract the disease related entities, and the relation extraction model is adopted to judge whether the recognized entity pairs have the semantic relation or not, so that the alignment problem of multi-source terms is solved, and a structured and reasonable industry knowledge base is formed; a complete logic chain is formed from a disease phenomenon to a maintenance decision, and the accuracy and recall rate of extracting entities and relationships in a complex unstructured text are improved.

Description

Technical Field

[0001] This invention relates to the field of transportation infrastructure maintenance technology, and in particular to a method for constructing a road defect knowledge graph and aligning entities. Background Technology

[0002] The field of highway pavement maintenance has accumulated a wealth of textual knowledge, such as the "Technical Specifications for Highway Asphalt Pavement Maintenance" and the "Standards for Highway Technical Condition Assessment," which define in detail the types of defects, grading standards, causal analysis, and corresponding maintenance measures. Engineering documents include test reports, test data documents, maintenance manuals, construction method descriptions, and expert diagnostic reports. Currently, this knowledge mainly exists in the form of "unstructured text" (Word / PDF), and in practical applications, it mainly relies on manual review or simple keyword searches (such as searching for "ruts" in PDFs). However, knowledge is often scattered across different specifications and chapters. Existing methods cannot understand the logical connections between these scattered texts, nor can they construct a complete logical chain from [disease type] → (based on) → [grading standards] → (analysis) → [causal mechanism] → (derivation) → [corresponding maintenance measures]. This results in the system being unable to perform full-link automated diagnosis, and can only provide fragmented information retrieval. Consequently, existing highway management systems cannot directly answer questions involving multi-hop reasoning, such as "why is this solution recommended" or "which level does this disease belong to," relying more on the experience of engineers or experts. Furthermore, different documents describe the same disease differently, such as "cracking" versus "net-like cracks" and "map-like cracks," lacking automated semantic alignment methods, leading to fragmented knowledge that is difficult to reuse. Summary of the Invention

[0003] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method for constructing and aligning road defect knowledge graphs.

[0004] The objective of this invention is achieved through the following technical solution: a method for constructing and aligning a road defect knowledge graph, comprising the following steps:

[0005] S1: Select data to construct an ontology model;

[0006] S2: Preprocess the data and use an entity recognition model to extract disease-related entities;

[0007] S3: Determine whether there is a semantic relationship between the identified entity pairs based on the relation extraction model;

[0008] S4: Align disease terms and entities;

[0009] S5: Store and apply knowledge graphs.

[0010] Preferably, in step S1, an open-source ontology editing tool is used to construct an ontology model. The five-dimensional ontology structure of the ontology model includes disease entities, disease characteristics, causal mechanisms, treatment measures, and assessment levels.

[0011] Preferably, in step S2, the data preprocessing specifically involves using the pdfplumber tool to parse the PDF document, removing headers and footers, and using the sliding window method to segment the long document into a short text sequence with a length not exceeding 512 tokens, with the window step size set to 256.

[0012] Preferably, in step S2, the entity recognition model includes a BERT embedding layer, a BiLSTM encoding layer, and a CRF decoding layer;

[0013] BERT embedding layers take the input text sequence The input is fed into an entity recognition model, which outputs a 768-dimensional dynamic word vector sequence corresponding to each character. ;

[0014] The BiLSTM coding layer will sequence Features are obtained by inputting into a bidirectional long short-term memory network (BiLSTM).

[0015] The CRF decoding layer is used to adjust the dependencies between tags.

[0016] Preferably, the Bidirectional Long Short-Term Memory (BiLSTM) network is used to compute the hidden states via a feedforward LSTM.

[0017] ;

[0018] Backward LSTM computes hidden states

[0019] ;

[0020] in, For the first The input feature vector of each character, Let be the hidden layer state vector at the current moment. Let be the hidden layer state vector from the previous time step. This is the hidden layer state vector for the next time step;

[0021] Then concatenate to obtain complete features .

[0022] Preferably, the CRF decoding layer decodes the input sequence. and predicted label sequence Calculate the score.

[0023] ;

[0024] in, The first output of BiLSTM The character was marked as The probability, For from the tag Transferred to The transition probability matrix;

[0025] Model training employs maximum likelihood estimation, minimizing the negative log-likelihood loss:

[0026] ;

[0027] Using the Viterbi algorithm to solve Maximum optimal path .

[0028] Preferably, in step S3, the relation extraction model includes an input layer and a segmented convolutional layer;

[0029] The input vector in the input layer is composed of word vectors and position vectors. The position vectors describe the path from the current word to the head entity. Tail-end entity The relative distance is mapped to a low-dimensional vector through a lookup table;

[0030] The segmented convolutional layer divides the sentence feature map into three segments based on the positions of the two entities: before the head entity, between the head and tail entities, and after the tail entity. Then, max pooling is performed on each of these three segments, as shown in the following formula:

[0031] ;

[0032] in, For the first The output value of each segment after max pooling These are the local features of the sentence extracted by the model through convolutional kernels. Segments into which a sentence is divided. Before the head entity, Between the head and tail entities, After the tail entity;

[0033] The three pooling results are concatenated to obtain .

[0034] Preferably, step S4 further includes the following step:

[0035] S41: Build an inverted index for the extracted entities and only retrieve entities with the same entity type as candidate pairs;

[0036] S42: For any two candidate entities and Calculate their overall similarity ,

[0037] ;

[0038] in, These are the weighting coefficients;

[0039] S43: Calculate literal similarity using normalized Levenshtein edit distance. ,

[0040] ;

[0041] in, To be Transform into The minimum number of insertions, deletions, or replacements required. The length of the string;

[0042] S44: Calculate semantic similarity ,

[0043] ;

[0044] in, For entities The high-dimensional semantic feature vector obtained by calculating the SBERT model; For entities The high-dimensional semantic feature vector obtained by calculating the SBERT model;

[0045] when When, it is determined that the two refer to the same concept, where, This is the alignment threshold.

[0046] The present invention has the following advantages: By constructing a disease ontology model, and then using an entity recognition model to extract disease-related entities and a relation extraction model to determine whether the identified entity pairs have semantic relationships, the present invention solves the alignment problem of multi-source terms, forming a structured and reasonable industry knowledge base, enabling a complete logical chain from disease phenomena to maintenance decisions, and improving the accuracy and recall rate of extracting entities and relations in complex unstructured text. Attached Figure Description

[0047] Figure 1 A schematic diagram illustrating the process of constructing and aligning entities for a road defect knowledge graph. Detailed Implementation

[0048] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0049] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0050] It should be noted that, unless otherwise specified, the embodiments and features described in this invention can be combined with each other.

[0051] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0052] In the description of this invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, or the orientation or positional relationship commonly used when the product of this invention is in use, or the orientation or positional relationship commonly understood by those skilled in the art. They are only used for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this invention. In addition, the terms "first," "second," etc., are only used to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0053] In the description of this invention, it should also be noted that, unless otherwise explicitly specified and limited, the terms "set," "install," "connect," and "link" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0054] In this embodiment, as Figure 1 As shown, a method for constructing and aligning a road defect knowledge graph includes the following steps:

[0055] S1: Select data to construct an ontology model; preferably, in step S1, an open-source ontology editing tool is used to construct the ontology model. The five-dimensional ontology structure of the ontology model includes the disease entity, disease characteristics, cause mechanism, treatment measures, and assessment level. Specifically, the data sources are selected from the "Specifications for Design of Asphalt Pavement Maintenance" (JTG 5421), "Standards for Assessment of Highway Technical Condition" (JTG 5210), "Technical Standards for Highway Maintenance" (JTG 5110-2023), and the supervision logs and inspection reports of actual maintenance projects. The OWL ontology was constructed using the open-source ontology editing tool Protégé. The specific five-dimensional ontology structure is defined as follows: Defect: Specifically refers to pavement damage phenomena, such as transverse cracks, alligator cracks, and potholes; Feature: Quantitative or qualitative description of the defect, such as crack width >3mm (quantitative) or network cracks (qualitative); Mechanism: The root cause of the defect, such as reflective cracks in the base course or asphalt aging; Measure: Includes Material (e.g., modified emulsified asphalt) and Technique (e.g., micro-surfacing); Grade: The severity level of the defect, such as mild, moderate, and severe.

[0056] S2: Preprocess the data and use an entity recognition model to extract disease-related entities;

[0057] S3: Determine whether there is a semantic relationship between the identified entity pairs based on the relation extraction model; preferably, in step S3, the relation extraction model includes an input layer and a segmented convolutional layer;

[0058] The input vector in the input layer is composed of word embeddings and position embeddings. The position embedding describes the path from the current word to the head entity. Tail-end entity The relative distance is mapped to a low-dimensional vector through a lookup table;

[0059] Before the segmented convolutional layer divides the sentence feature map into head entities based on the positions of the two entities,... Between the head and tail entities After the tail entity Then, perform max pooling on these three segments separately, as shown in the following formula:

[0060] ;

[0061] in, For the first The output value of each segment after max pooling These are the local features of the sentence extracted by the model through convolutional kernels. Segments into which a sentence is divided. Before the head entity, Between the head and tail entities, After the tail entity;

[0062] The three pooling results are concatenated to obtain This allows for the capture of structural information between entities while preserving features. For tabular data with a highly fixed format, a regular expression-based matching method can also be used.

[0063] S4: Align disease terms and entities;

[0064] S5: The knowledge graph is stored and applied. Specifically, it is stored and managed using the Neo4j graph database, where nodes correspond to entities, labels represent entity types, and edges correspond to semantic relationships. Alternatively, RDF (Resource Description Framework) triples can be used for storage. Applications can implement multi-hop reasoning based on the Cypher query language. By constructing a disease ontology model, and then using an entity recognition model to extract disease-related entities and a relationship extraction model to determine whether semantic relationships exist between identified entity pairs, the alignment problem of multi-source terms is solved, forming a structured, reasonable industry knowledge base. This creates a complete logical chain from disease phenomena to maintenance decisions, improving the accuracy and recall of extracting entities and relationships from complex unstructured text.

[0065] Furthermore, in step S2, the data preprocessing specifically involves using the pdfplumber tool to parse the PDF document, removing headers and footers, and using the sliding window method to segment the long document into short text sequences with a length not exceeding 512 tokens, with the window step size set to 256 to preserve the semantic integrity of the context; the BIOES (Begin, Inside, End, Single) annotation method is used, for example, the severe rutting of asphalt pavement is annotated as [O, O, O, O, O, O, B-Grade, E-Grade, B-Defect, E-Defect].

[0066] Furthermore, in step S2, the entity recognition model includes a BERT embedding layer, a BiLSTM encoding layer, and a CRF decoding layer;

[0067] BERT embedding layers take the input text sequence The input is fed into an entity recognition model, which outputs a 768-dimensional dynamic word vector sequence corresponding to each character. ;

[0068] The BiLSTM coding layer will sequence Features are obtained by inputting into a Bidirectional Long Short-Term Memory (BiLSTM) network; specifically, the BiLSTM network uses a forward LSTM to compute the hidden states.

[0069] ;

[0070] Backward LSTM computes hidden states

[0071] ;

[0072] in, For the first The input feature vector of each character, Let be the hidden layer state vector at the current moment. Let be the hidden layer state vector from the previous time step. This is the hidden layer state vector for the next time step;

[0073] Then concatenate to obtain complete features .

[0074] The CRF decoding layer is used to adjust the dependencies between tags. Specifically, the CRF decoding layer processes the input sequence... and predicted label sequence Calculate the score.

[0075] ;

[0076] in, The first output of BiLSTM The character was marked as The probability, For from the tag Transferred to The transition probability matrix;

[0077] Model training employs maximum likelihood estimation, minimizing the negative log-likelihood loss:

[0078] ;

[0079] Using the Viterbi algorithm to solve Maximum optimal path Specifically, compared to traditional text storage, steps S1 to S3 of this invention transform the text scattered across various chapters of the specification into a structured graph, clearly demonstrating the complete decision-making logic from disease grading to cause analysis and then to treatment plan.

[0080] In this embodiment, step S4 further includes the following step:

[0081] S41: Construct an inverted index for the extracted entities, and only retrieve entities with the same entity type as candidate pairs to reduce computational complexity;

[0082] S42: For any two candidate entities and Calculate their overall similarity ,

[0083] ;

[0084] in, This is the weighting coefficient, with a value of 0.4;

[0085] S43: Calculate literal similarity using normalized Levenshtein edit distance. ,

[0086] ;

[0087] in, To be Transform into The minimum number of insertions, deletions, or replacements required. Given the string length, this indicator can effectively identify literal variations such as "severe ruts" and "ruts (severe)".

[0088] S44: Calculate semantic similarity In other words, the Sentence-BERT (SBERT) model is used to generate semantic vectors for entities. SBERT employs a Siamese Network structure, shares BERT parameters, and directly outputs sentence vectors with semantic information. and This indicator can effectively identify synonyms such as "cracking" and "netting," which have different literal meanings but the same semantic meaning.

[0089] ;

[0090] in, For entities The high-dimensional semantic feature vector obtained by calculating the SBERT model; For entities The high-dimensional semantic feature vector obtained by calculating the SBERT model;

[0091] when When the two refer to the same concept, it is determined that they are referring to the same concept. The alignment threshold is 0.85.

[0092] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for constructing and aligning a road defect knowledge graph, characterized in that: Includes the following steps: S1: Select data to construct an ontology model; S2: Preprocess the data and use an entity recognition model to extract disease-related entities; S3: Determine whether there is a semantic relationship between the identified entity pairs based on the relation extraction model; S4: Align disease terms and entities; S5: Store and apply knowledge graphs.

2. The method for constructing and aligning road defect knowledge graphs according to claim 1, characterized in that: In step S1, an ontology model is constructed using an open-source ontology editing tool. The five-dimensional ontology structure of the ontology model includes disease entities, disease characteristics, causal mechanisms, treatment measures, and assessment levels.

3. The method for constructing and aligning road defect knowledge graphs according to claim 2, characterized in that: In step S2, the data preprocessing specifically involves using the pdfplumber tool to parse the PDF document, removing headers and footers, and using the sliding window method to segment the long document into a short text sequence with a length not exceeding 512 tokens, with the window step size set to 256.

4. The method for constructing and aligning road defect knowledge graphs according to claim 3, characterized in that: In step S2, the entity recognition model includes a BERT embedding layer, a BiLSTM encoding layer, and a CRF decoding layer. The BERT embedding layer takes the input text sequence The input is fed into an entity recognition model, which outputs a 768-dimensional dynamic word vector sequence corresponding to each character. ; The BiLSTM coding layer will sequence Features are obtained by inputting into a bidirectional long short-term memory network (BiLSTM). The CRF decoding layer is used to adjust the dependencies between tags.

5. The method for constructing and aligning road defect knowledge graphs according to claim 4, characterized in that: Bidirectional Long Short-Term Memory (BiLSTM) network uses a feedforward LSTM to compute hidden states. ； Backward LSTM computes hidden states ； in, For the first The input feature vector of each character, Let be the hidden layer state vector at the current moment. Let be the hidden layer state vector from the previous time step. This is the hidden layer state vector for the next time step; Then concatenate to obtain complete features .

6. The method for constructing and aligning road defect knowledge graphs according to claim 5, characterized in that: The CRF decoding layer processes the input sequence. and predicted label sequence Calculate the score. ； in, The first output of BiLSTM The character was marked as The probability, For from the label Transferred to The transition probability matrix; Model training employs maximum likelihood estimation, minimizing the negative log-likelihood loss: ； Using the Viterbi algorithm to solve Maximum optimal path .

7. The method for constructing and aligning road defect knowledge graphs according to claim 6, characterized in that: In step S3, the relation extraction model includes an input layer and a segmented convolutional layer; The input vector in the input layer is composed of word vectors and position vectors. The position vectors describe the path from the current word to the head entity. Tail-end entity The relative distance is mapped to a low-dimensional vector through a lookup table; The segmented convolutional layer divides the sentence feature map into three segments based on the positions of the two entities: before the head entity, between the head and tail entities, and after the tail entity. Then, max pooling is performed on each of these three segments, as shown in the following formula: ； in, For the first The output value of each segment after max pooling These are the local features of the sentence extracted by the model through convolutional kernels. Segments into which a sentence is divided. Before the head entity, Between the head and tail entities, After the tail entity; The three pooling results are concatenated to obtain .

8. The method for constructing and aligning road defect knowledge graphs according to claim 7, characterized in that: Step S4 also includes the following steps: S41: Build an inverted index for the extracted entities and only retrieve entities with the same entity type as candidate pairs; S42: For any two candidate entities and Calculate their overall similarity , ； in, These are the weighting coefficients; S43: Calculate literal similarity using normalized Levenshtein edit distance. , ； in, To be Transform into The minimum number of insertions, deletions, or replacements required. The length of the string; S44: Calculate semantic similarity , ； in, For entities The high-dimensional semantic feature vector obtained by calculating the SBERT model; For entities The high-dimensional semantic feature vector obtained by calculating the SBERT model; when When, it is determined that the two refer to the same concept, where, This is the alignment threshold.

Images