A knowledge graph-based method for analyzing space launch supportability conditions

Abstract

The present invention provides a method for analyzing the supportability conditions of aerospace measurement and launch based on a knowledge graph, comprising: constructing a knowledge graph in the field of aerospace measurement and launch, representing the measurement and launch knowledge in the form of entities and relationships, and enabling semantic search of knowledge through the knowledge graph in the field of measurement and launch; establishing a supportability analysis workflow, and pushing knowledge to each step in the workflow based on the knowledge graph in the field of aerospace measurement and launch, thereby implementing supportability condition analysis of the measurement and launch process from the perspective of multi-professional collaboration; and utilizing the associations between entity nodes in the knowledge graph to push relevant auxiliary reference knowledge in the supportability condition analysis work to professional personnel. The present invention improves the efficiency and accuracy of supportability condition analysis work by constructing a knowledge graph in the field of aerospace measurement and launch, and gradually pushing relevant auxiliary reference knowledge to professional personnel according to the workflow.

Description

Technical Field

[0001] The present invention relates to the field of knowledge graphs, and in particular to a method for analyzing space launch supportability conditions based on knowledge graphs. Background Art

[0002] In recent decades, my country's aerospace technology has achieved rapid development, significantly enhancing its space launch capabilities. Simultaneously, the demand for using information technology to evaluate and assist in the guidance of space launch missions has also increased, requiring more efficient and rational analysis of launch process flows.

[0003] Supportability analysis is a crucial auxiliary task in the launch and test process. This involves analyzing and assessing supportability requirements for the launch and test process, adhering to the launch site's technical specifications for spacecraft and carrier vehicles. This involves analyzing supportability factors specific to each launch mission, including water, heating, electricity, non-standard equipment, and refueling and gas supply. This allows for the identification of potential inter-system issues before actual joint training missions, thereby improving mission reliability. The volume of data generated by the launch and test process regarding various supportability requirements is substantial and scattered across business systems in electronic or paper form, hindering the ability to mine this data for value. Traditional databases store data in a table structure, with tables linked only by primary and foreign keys. This makes fast multi-hop queries impossible when dealing with large amounts of data. Existing knowledge push technologies, particularly those based on collaborative filtering, suffer from the sparseness of user-item interaction data and cold start issues, leading to inaccurate push results.

[0004] The knowledge graph, which combines relevant research results in the field of artificial intelligence, focuses more on representing knowledge in a structured form. By modeling entities, relationships, and attributes, it can handle complex and diverse association analyses, infer unknown implicit relationships, and play an important role in the analysis of space launch support conditions. Summary of the Invention

[0005] The purpose of the present invention is to provide a method for analyzing the support conditions of aerospace measurement and launch based on a knowledge graph. By constructing a knowledge graph in the field of aerospace measurement and launch, relevant auxiliary reference knowledge is gradually pushed to professionals according to the work flow, which can effectively improve the efficiency and accuracy of the support condition analysis work.

[0006] The technical solution to achieve the purpose of the present invention is: a method for analyzing space launch supportability conditions based on a knowledge graph, comprising the following steps:

[0007] Step 1: Build a knowledge graph in the field of aerospace testing and launch;

[0008] Step 2: Semantic search based on the knowledge graph in the field of aerospace measurement and launch;

[0009] Step 3: Establish a supportability condition analysis workflow;

[0010] Step 4: Push knowledge based on the assurance condition analysis of the test and development domain knowledge graph;

[0011] Step 5: If the supportability condition analysis work has been completed, terminate the push; otherwise, repeat steps 3 to 4.

[0012] Preferably, constructing a knowledge graph in the field of aerospace measurement and launch includes the following steps:

[0013] Step 1.1: Clarify the concept and scope of the test and development domain ontology and collect corpus data in the test and development domain;

[0014] Step 1.2: Perform entity recognition and relationship extraction on the corpus data in the test and development field to obtain the triple representation of knowledge.

[0015] The first step is entity recognition. This method proposes an entity recognition method that integrates multi-source dictionary information. This method combines the advantages of both general domain dictionaries and test and hair dictionaries. First, a character sequence is mapped to a corresponding character vector sequence through a pre-trained model. Then, the same method is used to obtain general dictionary information and test and hair dictionaries. The dictionary information and the test and hair dictionaries are integrated to obtain multi-source dictionary information, which is of great help in entity recognition in test and hair corpora.

[0016] The second step is relation extraction. This paper proposes a conditional relation extraction method that combines entity weight information. The entities in the test and development domain corpus data are divided into two categories according to their weights. Different pooling methods are applied to the vector sequences output by the BERT preprocessing model, avoiding the use of the same pooling method for all entities and better expressing the entity semantic information.

[0017] Step 1.3: Store the triples obtained in Step 1.2 in a graph database to complete the construction of the aerospace test and launch domain knowledge graph. This method uses the Neo4j graph database to store the extracted triples and visualize the test and launch domain knowledge graph.

[0018] Preferably, semantic search based on the knowledge graph in the field of aerospace measurement and launch includes the following steps:

[0019] Step 2.1: Perform entity recognition and relationship extraction on the query sentences in the supportability condition analysis work to obtain question triples;

[0020] Step 2.2: Align the entities in the question triples with the entities with the same semantics in the test domain knowledge graph;

[0021] Step 2.3: Match the question triples with the test domain knowledge graph to obtain search results.

[0022] Preferably, establishing a supportability condition analysis workflow includes the following steps:

[0023] This method uses the BPMN2.0 specification to describe the assurance condition analysis workflow. The basic elements used to describe the workflow include sequence flow, parallel gateway, event, sub-process and step description.

[0024] Preferably, the push of assurance condition analysis knowledge based on the test and launch domain knowledge graph includes the following steps:

[0025] Step 3.1: Extract entities from the contextual information of the supportability condition analysis workflow established by the personnel;

[0026] Step 3.2: Use the graph embedding method of the graph attention network to vectorize each node in the test and development domain knowledge graph, project all nodes into a low-dimensional dense vector space, and map the entity extraction results of step 3.1 into a set of process vectors;

[0027] Step 3.3: Calculate the cosine similarity and set the similarity threshold to obtain nodes with similar semantics to the elements in the process vector set, and use the inverted index method to recall candidate texts;

[0028] Step 3.4: Calculate the text matching degree of each element in the candidate text set with the context of the assurance condition analysis workflow, and push the top-K results of the final calculation results to the post personnel.

[0029] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) Compared with the traditional supportability condition analysis work that requires reference to relevant documents or query databases, the present invention can realize semantic search of knowledge based on the knowledge graph of the measurement and launch field, making the query process more intuitive and flexible, and the query efficiency is also higher; (2) Compared with the traditional supportability condition analysis, the work content of various professional positions is complex and highly dependent on professional skills, knowledge and experience accumulation. The present invention can use the association relationship between entity nodes in the knowledge graph structure to push relevant auxiliary reference knowledge in the supportability condition analysis work to professional position personnel; (3) Compared with the traditional supportability condition analysis that is only carried out for a single equipment, the present invention establishes a supportability analysis workflow and pushes knowledge to each step in the workflow based on the knowledge graph of the aerospace measurement and launch field, thereby realizing the supportability condition analysis work of the measurement and launch process from the perspective of multi-professional collaboration. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 This is the overall process of the method of the present invention.

[0031] Figure 2 This is a flow chart of entity recognition in the method of the present invention.

[0032] Figure 3 This is a flowchart of the relationship extraction method of the present invention.

[0033] Figure 4 This is a flow chart for pushing knowledge based on the assurance condition analysis knowledge graph in the test and development field in the method of the present invention. DETAILED DESCRIPTION

[0034] The knowledge graph contains a large amount of information about entities and the relationships between them, which can serve as effective auxiliary information to enrich the descriptions of users and items, improve the accuracy of recommendation algorithms, and alleviate and solve problems in background technologies. This embodiment provides a method for analyzing aerospace test and launch supportability conditions based on a knowledge graph, including: constructing a knowledge graph in the field of aerospace test and launch, representing test and launch knowledge in the form of entities and relationships. Compared with the need to refer to relevant documents or query databases in traditional supportability condition analysis, the knowledge graph in the field of test and launch can realize semantic search of knowledge, making the query process more intuitive and flexible, and improving query efficiency; a method for pushing knowledge for supportability condition analysis based on the knowledge graph in the field of aerospace test and launch. Compared with the traditional supportability condition analysis in which the work content of various professional positions is complex and highly dependent on professional skills, knowledge and experience accumulation, the present invention can utilize the association relationship between entity nodes in the knowledge graph structure to push relevant auxiliary reference knowledge in the supportability condition analysis work to professional position personnel; establishing a supportability condition analysis workflow. Compared with traditional supportability condition analysis that is only performed for a single piece of equipment, the present invention establishes a supportability analysis workflow and pushes knowledge for each step in the workflow based on the knowledge graph in the field of aerospace test and launch, thereby realizing supportability condition analysis work for the test and launch process from the perspective of multi-professional collaboration.

[0035] Combine Figure 1 A method for analyzing space launch supportability conditions based on knowledge graphs includes the following steps:

[0036] Step 1: Build a knowledge graph in the field of aerospace testing and launch;

[0037] Step 2: Perform semantic search based on the knowledge graph in the field of aerospace measurement and launch;

[0038] Step 3: Establish a supportability condition analysis workflow;

[0039] Step 4: Push knowledge based on the assurance condition analysis of the test and development domain knowledge graph;

[0040] Step 5: If the supportability condition analysis work has been completed, terminate the push; otherwise, repeat steps 3 to 4.

[0041] Constructing the aerospace test and launch domain knowledge graph in step 1 includes the following steps:

[0042] Step 1.1: Clarify the concept and scope of the measurement and launch domain ontology, and collect corpus data in the measurement and launch domain. The present invention aims at constructing the ontology concept by summarizing and generalizing the existing support condition domain ontology, analyzing the data sources such as the launch site technical manual, and consulting relevant technical personnel and domain experts, and setting up a 4-level ontology tree. The first-level ontology includes 4 major categories: measurement and launch procedures, technical workshops, launch stations, and equipment specialties. The second-level to fourth-level ontologies are specific divisions of the previous level. In addition, the "hasDev" relationship between the workshop / room and the equipment, the "belong" relationship between the specialty / category and the category / equipment, the "hasAttr" relationship between the equipment and the attribute, the "is" relationship between the attribute and the attribute value, and other relationships are designed;

[0043] Step 1.2: Perform entity recognition and relationship extraction on the corpus data in the test and development field to obtain the triple representation of knowledge.

[0044] The first step is entity recognition. This method proposes an entity recognition method that integrates multi-source dictionary information, such as Figure 2 As shown in the figure, this method has the advantages of both general domain dictionaries and test and hair domain dictionaries. First, the character sequence is mapped to the corresponding character vector sequence through the pre-training model. Then, the same method is used to obtain the general dictionary information and the test and hair domain dictionary. The dictionary information and the test and hair domain dictionary are fused to obtain multi-source dictionary information, which is of great help to the entity recognition work of the test and hair domain corpus. The following mainly introduces the method of multi-source dictionary information fusion. First, each sentence s in the test and hair domain corpus data is converted into a character sequence {c1, c2, ..., c n}, where n is the length of sentence s, and each character c is obtained by the BERT model i Vector representation of As shown below.

[0045]

[0046] Using the general dictionary D g After obtaining all matching words, the matching words are grouped and annotated using the BMES annotation scheme. After the annotation is completed, B(c i )、M(c i )、E(c i ) and S(c i ) Four word sets. If there are still empty word sets after the entire character sequence is marked, add "None" to the word set, as shown below.

[0047]

[0048]

[0049]

[0050]

[0051] Among them, w i,k For character sequences starting with c i Start c k The ending word, w j,k and w j,i The meaning is the same.

[0052] In order to avoid the situation where the number of words in the same word set is different, the word set S needs to be compressed. Considering that different words have different weights in entity recognition work, this method uses a weighted method to compress the word set, and the compressed vector of the word set is obtained by the following formula.

[0053]

[0054]

[0055] Where z(w) is the word frequency of w in the corpus data, v P is a weighting function, and P includes four BMES labels.

[0056] At this point, the vector representations of the four word sets can be combined to obtain the universal dictionary feature H g , as shown below.

[0057] H g =[v B (c i );v M (c i );v E (c i );v S (c i )]

[0058] By repeating the above steps, we can also obtain the dictionary features H in the test field b , the character vector representation of the information of the fused multi-source dictionary is obtained by the following formula.

[0059] x c =[x c ;H g ;H b ]

[0060] The second step is relation extraction. This invention proposes a guarantee condition relation extraction method combining entity weight information, such as Figure 3As shown in the figure, entities in the test and development domain corpus are divided into two categories based on weights. Different pooling methods are applied to the vector sequences output by the BERT preprocessing model to avoid using the same pooling method for all entities, thereby better expressing the entity semantic information. The following mainly introduces the classification pooling operation.

[0061] First, the TF-IDF method is used to calculate the weight of the entity by counting the word frequency and the inverse document frequency, and then the weight threshold θ is set. For the entity weight w in the sentence e , when w e <θ, this entity is regarded as a relatively unimportant entity, and the vector sequence corresponding to the entity output by BERT is processed using average pooling. e When ≥θ, the entity is considered relatively important, and the vector sequence corresponding to the entity is max-pooled. The pooled vector is then fed into the activation function for activation. The entity vector representation is calculated as follows.

[0062]

[0063] Among them, i and j represent the starting position and ending position of the entity in the sentence. Through i and j, we can locate the specific entity word and get the weight of this entity word. Tanh is the hyperbolic tangent function. [t i ,t(i+1),…,t j ] is the sentence representation vector output by the BERT preprocessing model, and weight_pooling is the classification pooling operation. The implementation process is as follows.

[0064]

[0065] Among them, w e Represents vector weight, meanpool is the average pooling method, and maxpool is the maximum pooling operation.

[0066] Step 1.3: Store the triples obtained in Step 1.2 in a graph database to complete the construction of the aerospace test and launch domain knowledge graph. This method uses the Neo4j graph database to store the extracted triples and visualize the test and launch domain knowledge graph.

[0067] The semantic search based on the aerospace test and launch domain knowledge graph in step 2 includes the following steps:

[0068] Step 2.1: Perform entity recognition and relationship extraction on the query sentences in the supportability condition analysis work to obtain question triples;

[0069] Step 2.2: Align the entities in the question triples with the entities with the same semantics in the test domain knowledge graph;

[0070] Step 2.3: Match the question triples with the test domain knowledge graph to obtain search results.

[0071] The establishment of the supportability condition analysis workflow in step 3 includes the following steps:

[0072] This method uses the BPMN 2.0 specification to describe the supportability analysis workflow. The basic elements used to describe the workflow include sequence flows, parallel gateways, events, sub-processes, and step descriptions. The following example illustrates a specific supportability analysis workflow. In the supportability analysis task, the docking and installation of the spacecraft / fairing assembly with the rocket in the rocket's vertical assembly and test facility is a crucial step. It is also a crucial step in the "three verticals and one remote" (vertical assembly, vertical testing, vertical transfer, and remote launch) test and launch model currently used by my country's launch vehicles, namely, "vertical assembly." This step is complex, involving numerous concurrent process steps and requiring collaboration between crane control personnel and ground support personnel. Modeling using the BPMN specification allows for an intuitive and rich description of this process and behavior, and allows for control operations such as forward and reverse flow.

[0073] Combine Figure 4 In step 4, the knowledge push of the assurance condition analysis based on the test and launch domain knowledge graph includes the following steps:

[0074] Step 4.1: Extract entities from the contextual information of the supportability condition analysis workflow established by the personnel;

[0075] Step 4.2: Use the graph embedding method of the graph attention network to vectorize each node in the knowledge graph of the test and development domain, project all nodes into a low-dimensional dense vector space, and map the entity extraction results of step 3.1 into a set of process vectors;

[0076] The first step is to represent the constructed knowledge graph in the testing and development field as a weighted adjacency matrix, using weights to distinguish the relationships between nodes.

[0077] In the second step, the graph attention network layer learns the relationship between any adjacent nodes x in the graph by introducing the attention weight matrix i and x j The weight of is used to determine whether adjacent nodes have a relationship based on the adjacency matrix. The adjacency matrix A is calculated as follows.

[0078] A=(a ij )+I N

[0079] Among them, a ij Represents the attention correlation coefficient matrix between node i and node j, I Nis the identity matrix.

[0080] The update mechanism of the graph attention network is as follows.

[0081]

[0082] in, and are the vector representations of the i-th node in the l+1th and l-th layers, respectively, W l is the parameter matrix of the lth layer, and σ is the nonlinear activation function. After the graph attention network layer updates the weights of each node in the graph, the output of this layer is an m×n feature matrix, where m represents the number of nodes in the graph and n represents the feature dimension.

[0083] Step 4.3: By calculating the cosine similarity and setting the similarity threshold, we can obtain nodes with similar semantics to the elements in the process vector set, and recall the candidate texts through the inverted index method; through the graph attention network embedding, we can allow the knowledge in the graph that is far away but has similar connection relationships and node features to be embedded in the low-dimensional adjacent vector space. At this time, the node distance in the vector space reflects the correlation between the knowledge. For example, the "vertical assembly hall bridge crane" is a device involved in the "spacecraft / fairing vertical assembly" process, so the two pieces of knowledge have similar connections in the vector space. Therefore, the vector distribution learned in a large amount of text will be in a similar space, so the candidate text set can be enriched by recalling the embedded vector nodes in the adjacent space;

[0084] Step 4.4: Calculate the textual matching degree of each element in the candidate text set against the context of the supportability analysis workflow, and send the top-K results of the final calculation to the personnel in charge. This method improves the DSSM text matching model by replacing the one-hot encoding used in the input layer of the original DSSM model with the BERT model and the DNN / CNN network used in the representation layer with the BiLSTM model. After obtaining the text semantic features, the matching layer finally calculates the cosine similarity of the two vectors to determine the textual matching degree between the candidate text and the user-established supportability analysis workflow context.

[0085] The above embodiments are illustrations of specific implementation methods of the present invention, rather than limitations of the present invention. Technicians in the relevant technical fields can make various changes and modifications to obtain corresponding equivalent technical solutions without departing from the spirit and scope of the present invention. Therefore, all equivalent technical solutions should be included in the patent protection scope of the present invention.

Claims

1. A method for analyzing space launch supportability conditions based on knowledge graph, characterized by: Including steps: Step 1: Construct a knowledge graph for aerospace test and launch, and represent the test and launch knowledge in the form of entities and relationships; Step 2: Perform semantic search based on the knowledge graph in the field of aerospace measurement and launch; Step 3: Establish a supportability condition analysis workflow; Step 4: Push knowledge based on the assurance condition analysis of the test and development domain knowledge graph; Step 5: If the supportability condition analysis has been completed, terminate the push; otherwise, repeat steps 3 to 4; The construction of the aerospace test and launch field knowledge graph specifically includes the following steps: Step 1.1: Clarify the concept and scope of the test and development domain ontology and collect corpus data in the test and development domain; Step 1.2: Perform entity recognition and relationship extraction on the corpus data in the test and development field to obtain a triple representation of knowledge; Step 1.3: Store the triples obtained in step 1.2 into the graph database to complete the construction of the knowledge graph in the field of aerospace measurement and launch. The entity recognition in step 1.2 adopts an entity recognition method that integrates multi-source dictionary information, specifically including: First, the character sequence is mapped to the corresponding character vector sequence through the BERT pre-training model. Then, the general dictionary information and the test and hair domain dictionary are obtained using the same method. The general dictionary information and the test and hair domain dictionary are fused to obtain the multi-source dictionary information. Each sentence s in the corpus data of the test and hair domain is converted into a character sequence {c1, c2, ..., c n }, where n is the length of sentence s, and each character c is obtained by the BERT pre-training model i Vector representation of for: Using the general dictionary D g After obtaining all matching words, the matching words are grouped and annotated using the BMES annotation scheme. After the annotation is completed, B(c i )、M(c i )、E(c i ) and S(c i ) Four word sets. If there are still empty word sets after the entire character sequence is marked, add "None" to the word set. The four word sets are: Among them, w i,k For character sequences starting with c i Start c k The ending word, w j,k and w j,i Similarly; Use the weighted method to compress the word set S and obtain the compressed vector of the word set: Where z(w) is the word frequency of w in the corpus data, v P is the weighting function, P includes four BMES labels; The vector representations of the four word sets are combined to obtain the universal dictionary feature H g : H g =[v B (c i );v M (c i );v E (c i );v S (c i )] Repeat the above steps to obtain the dictionary features H in the test field b , we get the character vector representation of the information of the fused multi-source dictionary: x c =[x c ;H g ;H b ]。 2. The method for analyzing space launch supportability conditions based on knowledge graph according to claim 1 is characterized in that: The relationship extraction in step 1.2 includes: dividing the entities in the test and development domain corpus data into two categories according to weights, and applying different pooling methods to the vector sequence output by the BERT preprocessing model. The different pooling methods specifically include: First, the word frequency and inverse document frequency are counted, and the TF-IDF method is used to calculate the weight of the entity. Then, the weight threshold θ is set. For the entity weight w in the sentence e , when w e <θ, this entity is regarded as a relatively unimportant entity, and the vector sequence corresponding to the entity output by BERT is processed using average pooling. e When ≥θ, this entity is regarded as a relatively important entity, the vector sequence corresponding to the entity is pooled using the maximum method, and the pooled vector is sent to the activation function for activation; the calculation method of the entity vector representation is: h e =tanh(weight_polling([t i ,t(i+1),…,t j ])) Among them, i and j represent the starting position and ending position of the entity in the sentence. The specific entity word is located by i and j to obtain the weight of the entity word. Tanh is the hyperbolic tangent function. [t i ,t(i+1),…,t j ] is the sentence representation vector output by the BERT preprocessing model, and weight_pooling is the classification pooling operation, specifically: Among them, w e Represents vector weight, meanpool is the average pooling method, and maxpool is the maximum pooling operation.

3. The method for analyzing space launch supportability conditions based on knowledge graph according to claim 1 is characterized in that: The triples in step 1.3 are stored in the Neo4j graph database.

4. The method for analyzing space launch supportability conditions based on knowledge graph according to claim 1 is characterized in that: The semantic search based on the knowledge graph in the field of aerospace measurement and launch specifically includes the following steps: Step 2.1: Perform entity recognition and relationship extraction on the query sentences in the supportability condition analysis work to obtain question triples; Step 2.2: Align the entities in the question triples with the entities with the same semantics in the test domain knowledge graph; Step 2.3: Match the question triples with the test domain knowledge graph to obtain search results.

5. The method for analyzing space launch supportability conditions based on knowledge graph according to claim 1 is characterized in that: The assurance condition analysis workflow is described using the BPMN 2.0 specification. The basic elements used to describe the workflow include sequence flow, parallel gateway, event, sub-process and step description.

6. The method for analyzing space launch supportability conditions based on knowledge graph according to claim 1 is characterized in that: The method for pushing supportability analysis knowledge based on the knowledge graph in the field of aerospace measurement and launch specifically includes the following steps: Step 4.1: Extract entities from the contextual information of the supportability condition analysis workflow established by the personnel; Step 4.2: Use the graph embedding method of the graph attention network to vectorize each node in the knowledge graph of the test and development domain, project all nodes into a low-dimensional dense vector space, and map the entity extraction results of step 4.1 into a set of process vectors; Step 4.3: Calculate the cosine similarity and set the similarity threshold to obtain nodes with similar semantics to the elements in the process vector set, and use the inverted index method to recall candidate texts; Step 4.4: Calculate the text matching degree of each element in the candidate text set with the context of the assurance condition analysis workflow, and push the top-K results of the final calculation results.

7. The method for analyzing space launch supportability conditions based on knowledge graph according to claim 6 is characterized in that: The step 4.2 specifically includes: 4.2.

1. Represent the constructed knowledge graph in the testing and development domain as a weighted adjacency matrix, using weights to distinguish the relationships between nodes. 4.2.2, the graph attention network layer introduces the attention weight matrix to learn any adjacent node x in the graph i and x j The weight of is used to determine whether adjacent nodes have a relationship based on the adjacency matrix. The adjacency matrix A is: A=(a ij )+I N Among them, a ij Represents the attention correlation coefficient matrix between node i and node j, I N is the identity matrix, and the update mechanism of the graph attention network is: in, and are the vector representations of the i-th node in the l+1th and l-th layers, respectively, W l is the parameter matrix of the lth layer, σ is the nonlinear activation function. After the graph attention network layer updates the weight of each node in the graph, the output of this layer is an m×n feature matrix, where m represents the number of nodes in the graph and n represents the feature dimension.

8. The method for analyzing space launch supportability conditions based on knowledge graph according to claim 6 is characterized in that: The text matching degree calculation adopts the DSSM text matching model, the input layer of the DSSM text matching model adopts the BERT model, and the representation layer adopts the BiLSTM model.

Images