A semantic retrieval method for unmanned aerial vehicle engine disassembly process knowledge graph

Abstract

The application discloses a semantic retrieval method for a UAV engine dismounting process knowledge graph, and relates to the technical field of semantic retrieval. i ∈M possible candidate entity set E i Noise deletion, through twice pruning, first, based on the entity mention category, the candidate entity is pruned for the first time, and then based on the surface similarity between the candidate entity and the entity mention, the second pruning is carried out; Target function construction, the collective entity linking target function of the UAV engine dismounting process knowledge graph is constructed; Target function solving, three models are established to solve the co-occurrence probability of the candidate entity group, the overall semantic similarity of the candidate entity group and the conditional probability between the entity mention and the candidate entity, which has the characteristics of high accuracy and high stability.

Description

Technical Field

[0001] This invention relates to the field of semantic retrieval technology, specifically to a semantic retrieval method for a knowledge graph of unmanned aerial vehicle (UAV) engine disassembly process. Background Technology

[0002] In recent years, knowledge graphs have been gradually applied in the field of health status management, including equipment fault tracing, fault diagnosis, and intelligent and scientific disassembly of equipment. Intelligent disassembly of unmanned equipment is a key technology for fault repair after the fault of unmanned engine is determined, and it is of great significance for supporting the rapid and safe disassembly of UAV engines. However, the current knowledge retrieval accuracy is limited and it is difficult to support the semantic retrieval of engine disassembly process knowledge, which makes it difficult to accurately query the disassembly process related to the disassembly target in the knowledge graph. Collective entity linking technology can perform knowledge retrieval more accurately by considering the contextual information of the disassembly target. However, there are still three prominent key problems in collective entity linking for UAV engine disassembly knowledge graphs that need to be solved: (1) the objective function solution of collective entity linking deviates from the idea of ​​"collective linking"; (2) the objective function solution does not highlight the strong logical relationship between engine disassembly process knowledge; and (3) the utilization rate of process knowledge in the knowledge graph is insufficient.

[0003] Entity linking technology is a key technology for applying the constructed knowledge graph to practical engineering, playing a decisive role in knowledge retrieval and recommendation. An entity linking task can be described as finding the unique entity in a knowledge base that corresponds to a keyword, completing the mapping process from keywords to the knowledge base, and achieving accurate knowledge retrieval.

[0004] However, most current entity linking methods are geared towards general domains. They typically input the retrieved keywords into third-party databases such as search engines to obtain keyword description pages, redirect pages, and links. This related knowledge is then used as supplementary information to complete entity links, improving their accuracy. Because search engines and other third-party databases store general knowledge, specialized knowledge specific to the UAV engine disassembly process is difficult to find in these databases. Therefore, these entity linking methods based on search engines and other third-party databases are highly unsuitable for highly specialized fields, resulting in low entity link accuracy and failure to meet usage conditions. Consequently, the information available for entity linking in the UAV engine disassembly process field is limited. Based on the above analysis, how to fully extract effective information from the current disassembly process knowledge graph to improve entity link accuracy is a current challenge. Summary of the Invention

[0005] This invention proposes a semantic retrieval method for knowledge graphs of UAV engine disassembly processes, which features high accuracy and stability.

[0006] The technical solution of the present invention is as follows:

[0007] A semantic retrieval method for a knowledge graph of unmanned aerial vehicle (UAV) engine disassembly process is proposed. This method operates on the text description of the actual engine disassembly process, uniquely locating the text description based on a given knowledge graph. The method involves entity mentions m. i The generated candidate entity set is in, Indicates entity mention m i The k-th candidate entity, ni represents the entity mentioning m. i The number of candidate entities,

[0008] Candidate entity generation: Find each entity mention m from a given knowledge graph. i The set of candidate entities E that may correspond to ∈M i ,

[0009] Noise removal is achieved through two pruning processes. First, candidate entities are pruned based on their mention categories. Then, a second pruning is performed based on the surface similarity between the candidate entities and their mentions.

[0010] Objective function construction: Constructing a collective entity linking objective function for a knowledge graph of UAV engine disassembly process.

[0011] To solve the objective function, three models were established to solve the co-occurrence probability of candidate entity groups, the overall semantic similarity of candidate entity groups, and the conditional probability between entity mentions and candidate entities.

[0012] As a further optimization of this scheme, the specific formula for constructing the objective function is as follows:

[0013]

[0014] Where m1, m2, ..., m q Given q entity mentions in a text search of a process;

[0015] Entity mentions m i Candidate entity set The ki-th candidate entity in;

[0016] E1×E2×…×E q The Cartesian product of all candidate entity sets mentioned by the entity;

[0017] (e1,e2,…,e q), entity mention group (m1, m2, ..., m q A candidate entity group;

[0018] P(m1,…,m q ;e1,…,e q ), entities e1, e2, ..., e q The probability of appearing in the same text segment;

[0019] P(e1,e2,…,e q ), entities e1, e2, ..., e q The probability of appearing in the same text segment;

[0020] P(m1,…,m q |e1,…,e q In a text, given candidate entities e1, e2, ..., e q When entities mention m1, m2, ..., m q The probability of occurrence.

[0021] As a further optimization of this scheme, P(m1,m2,…,m) is accurately estimated based on the overall semantic similarity between entity mention groups and candidate entity groups. q |e1,e2,…,e q The specific formula is as follows:

[0022]

[0023] Where |M| represents the number of entity references in a text.

[0024] Se(*), measures entity mention groups (m1, m2, ..., m q ) and candidate entity groups (e1, e2, ..., e q The overall semantic similarity function;

[0025] P(m i |e i Given candidate entity e i Below, the entity mentions m i The probability of occurrence

[0026] The entity link objective function is expressed as follows:

[0027]

[0028] As a further optimization of this scheme, the overall relevance of candidate entity groups is used to estimate the co-occurrence probability of candidate entity groups, and an entity relevance measurement model based on three 2-hop paths is established.

[0029]

[0030]

[0031] Among them, R(e i ,e j ), process entity e i and e j The correlation measurement function;

[0032] (e i ,r,e j In the knowledge graph, the process entity e i and e j There is a relationship between them;

[0033] R2(e i ,e j ), a correlation measurement function based on 2-hop paths;

[0034] Process entity e i to e j The path length, in this section

[0035] They are respectively process entities e i The number of process entities pointed to and e j The number of process entities it points to;

[0036] Representing process entity e respectively i With e j The number of first-type 2-hop paths and the number of second-type 2-hop paths;

[0037] Process entity e i to e j The third type of 2-hop path quantity;

[0038] β is a hyperparameter that measures the importance of a directed 2-hop path to other 2-hop paths.

[0039] As a further optimization of this scheme, a model for measuring the overall correlation of candidate entity groups based on strongly correlated sequences is constructed.

[0040]

[0041] Where Tree1(e1,e2,…,e q ), Entities e1, e2, ..., e are obtained based on the relevance of entity pairs. q Strongly correlated sequences;

[0042] (ei ,e j An edge in a strongly correlated sequence;

[0043] R(e i ,e j ), entity e i ,e j The correlation.

[0044] As a further optimization of this scheme, the solution for the overall semantic similarity of candidate entity groups includes quantifying the semantic similarity between any two process entities using cosine similarity or Euclidean distance based on the semantic vectors of the process entities.

[0045]

[0046] Process entity e i semantic vector;

[0047] Calculation process entity and process entity Semantic similarity functions, such as cosine similarity function and the reciprocal of Euclidean distance.

[0048] As a further optimization of this scheme, strongly correlated sequences are used to calculate the overall semantic similarity of candidate entity groups to highlight the strong logical connections between process knowledge.

[0049]

[0050] Where Tree2(e1,e2,…,e q ), Entities e1, e2, ..., e are obtained based on the semantic similarity of entity pairs. q Strongly correlated sequences.

[0051] As a further optimization of this scheme, edit distance is used to calculate entity mention m. i and candidate entity e i The surface similarity is used to estimate the conditional probability between entity mentions and candidate entities.

[0052]

[0053] Where, |m i |、|e i |, respectively, entity mentions m i and candidate entity e i The number of characters;

[0054] lev(m i ,e i ), entity mentions mi and candidate entity e i Edit distance;

[0055] Lv(m i ,e i ), body mentions m i and candidate entity e i Surface similarity.

[0056] As a further optimization of this scheme, a hyperparameter γ greater than 0 is added when calculating the surface similarity between entity mentions and candidate entities.

[0057] Attached Figure Description

[0058] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0059] Figure 1 This is an example of entity linking in the knowledge graph of the JH70 engine disassembly process in a specific embodiment;

[0060] Figure 2 These are the three 2-hop paths in the knowledge graph of the JH70 engine disassembly process in a specific embodiment;

[0061] Figure 3 Here is an example of a fourth-order correlation plot and its correlation sequence;

[0062] Figure 4 This section compares the entity linking precision, recall, and F1 score in a specific embodiment. Detailed Implementation

[0063] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0064] According to the fine-grained collective entity linking method for the knowledge graph of UAV engine disassembly process determined by the present invention, knowledge retrieval on the knowledge graph of UAV engine disassembly process can be completed according to the following steps:

[0065] Practical process knowledge retrieval and analysis

[0066] Retrieving actual process knowledge involves both fuzzy matching and semantic matching. Process engineers vary in their professional skills and prior knowledge. Therefore, when using a knowledge graph database of UAV engine disassembly processes for process knowledge retrieval, the input search text lacks a unified and standardized format. For example, in a real search, a technician might want to query the disassembly process related to "stator assembly." The actual input might be semantically related terms such as "stator assembly" or "stator final assembly."

[0067] Definition 1. Entity reference: An entity reference is the act of mentioning or referencing a specific entity (part, function, date, or tool used, etc.) in a text. For example: Figure 1 In the text search of the actual disassembly process of the JH70 engine, the bolded words such as "both hands", "clutch connecting sleeve", "hand", and "shaft sleeve-clutch" are all entity references.

[0068] Definition 2. Candidate Entity: An entity in the knowledge graph that may represent the same thing as the entity mentioned. For example: Figure 1 The "shaft sleeve-clutch" in the diagram likely represents the same thing as 65 other entities in the JH70 engine disassembly process knowledge graph, such as "oil pipe [clutch]" and "sleeve [clutch]". Therefore, all 65 entities are candidate entities for "shaft sleeve-clutch".

[0069] Definition 3. Entity Link: A set of entity mentions M = {m1, m2, ..., mn} given a text. q} and a knowledge graph G = (V, R, E). The goal of entity linking is to mention m for any entity in M. i ∈M, find the unique relation to m in the knowledge graph G. i Refers to the same entity v∈V.

[0070] Definition 4. Entity mention set, candidate entity set: Given a text, the entity mention set M = {m1, m2, ..., m} q}, and mentioning entities m through candidate entity generation methods. i The generated candidate entity set is in, Indicates entity mention m i The k-th candidate entity, ni represents the entity mentioning m. i The number of candidate entities. During entity linking, all entity mentions (m1, m2, ..., m...) q A group of entity mentions is an element in the Cartesian product of all candidate entity sets. This is referred to as the candidate entity group.

[0071] Definition 5. Collective Entity Links: Given a set of entity mentions M = {m1, m2, ..., mn} of a text. q} and a knowledge graph G = (V, R, E). Collective entity linking treats all entity mentions in this text as a whole, i.e., an entity mention group (m1, m2, ..., m...). q The goal is to find a set of references to entity mentions (m1, m2, ..., m) in a knowledge graph G. q ) refers to a group of candidate entities (e1, e2, ..., e) that represent the same real-world entity. q ), where e1, e2, ..., e q ∈V.

[0072] Candidate entity generation and two-step pruning strategy

[0073] Entity linking comprises two parts: candidate entity generation and candidate entity ranking. Candidate entity generation generates potential candidates for entity mentions, while candidate entity ranking finds the candidates most likely to represent the same semantic meaning as the entity mentions.

[0074] Candidate entity generation finds the set of possible candidate entities for each entity mention from a given knowledge graph. If the entity linked to an entity mention is not in its candidate entity set, the entity mention cannot be linked to the correct entity. Therefore, candidate entity generation should ensure, as far as possible, that the entity corresponding to the entity mention is in the candidate entity set, guaranteeing the recall rate of entity mentions. In reality, process engineers already have some prior knowledge related to the search target when retrieving disassembly process knowledge. Therefore, the search text entered by disassembly process engineers usually contains some entity mentions that are related to the search target or have similar appearances. For example, when the search target is "engines in the aviation field", disassembly process engineers may directly enter "aircraft engine", "engine", "engines for the aviation field", etc. Therefore, the following strategy is used to generate a set of candidate entities for each entity mention m. i ∈M generates a candidate entity set E i :

[0075] Entities in a knowledge graph.

[0076] After word segmentation of entities and entity mentions in the disassembly process knowledge graph, the entities and entity mentions have common words.

[0077] The entities in the disassembly process knowledge graph are abbreviations of entity references or entity references are abbreviations of entities.

[0078] Collective entity linking techniques typically have high time complexity. The generation of candidate entities can lead to a large number of irrelevant candidates, significantly increasing the computational complexity of the collective entity linking model. For example, the candidate entity for "sleeve used for clutch" has 65 candidates. To reduce the computational complexity of entity linking, two strategies are employed to prune the candidate entity set based on the category information of entity mentions and candidate entities, yielding the top-k candidate entities. First, candidate entities are pruned based on the entity mention category; second, a second pruning is performed based on the surface similarity between candidate entities and entity mentions. This pruning operation significantly narrows the solution scope for entity linking; for example, after pruning, the candidate entity for "sleeve used for clutch" becomes only 2.

[0079] Delete entities from the candidate entity set that do not belong to the category mentioned by the entity.

[0080] If the size of the candidate entity set is still greater than k, then calculate the surface similarity between the entity mention and the candidate entity, and set the similarity between the candidate entity's abbreviation and the candidate entity to 1. Finally, select the k candidate entities that are most similar to the entity mention.

[0081] Construction of a collective entity linking objective function for a knowledge graph of drone engine disassembly process

[0082] Given a set of entity mentions M = {m1, m2, ..., m} from a search text. q}, Entity reference type set T = {t1, t2, ..., t q} and a set of candidate entities with category information in, Indicates entity mention m i The corresponding candidate entity set, Indicates candidate entities The type is t i1 After pruning, the candidate entity set is pruned to E = {E1, E2, ..., E...} q},in, For entity mention m i The pruned candidate entity set, n i This represents the number of candidate entities after pruning. The goal of collective entity linking is to link entity mention groups (m1, m2, ..., m...). q In the knowledge graph, find a group of candidate entities that refer to the same thing as the entity mentioned. in, Specifically, as shown in formula (1):

[0083]

[0084] In the formula, m1, m2, ..., m q—Given q entity mentions in a text search for a process;

[0085] —Entity mention m i candidate entity set The ki-th candidate entity in;

[0086] E1×E2×…×E q —The Cartesian product of all candidate entity sets mentioned by the entity;

[0087] (e1,e2,…,e q — Entity mention group (m1, m2, ..., m) q A candidate entity group;

[0088] P(m1,…,m q ;e1,…,e q — Entity mention group (m1, m2, ..., m) q Link to candidate entity group (e1, e2, ..., e q The probability of ).

[0089] P(e1,e2,…,e q )——Entities e1, e2, ..., e q The probability of appearing in the same text segment;

[0090] P(m1,…,m q |e1,…,e q — In a text, given candidate entities e1, e2, ..., e q When entities mention m1, m2, ..., m q The probability of occurrence.

[0091] In a text retrieved about the disassembly process of a drone engine, the semantic relationships between the various entity mentions are very strong. Therefore, the entity mention group (m1, m2, ..., m...) q The stronger the overall semantic similarity of candidate entity groups (e1, e2, ..., e...), the better. q The stronger the overall semantic similarity of P(m1,m2,…,m) q |e1,e2,…,e q The larger the value, the greater the value. Furthermore, given candidate entity e... i Entity mentions m i The higher the probability of occurrence, the better P(m1,m2,…,m) becomes. q |e1,e2,…,e q The larger the value, the better. Based on the above analysis, P(m1,m2,…,m) is accurately estimated based on the overall semantic similarity between entity mention groups and candidate entity groups. q |e1,e2,…,eq As shown in formula (2).

[0092]

[0093] In the formula, |M| represents the number of entity references in a text.

[0094] Se(*) — Measures entity mention groups (m1, m2, ..., m q ) and candidate entity groups (e1, e2, ..., e q The overall semantic similarity function;

[0095] P(m i |e i — Given candidate entity e i Below, the entity mentions m i The probability of occurrence.

[0096] Furthermore, the objective function for entity linking can be expressed as formula (3).

[0097]

[0098] (1) Objective function solution – a fine-grained collective entity linking method based on strongly correlated sequences

[0099] To accurately solve the objective function and improve the entity linking accuracy of the disassembly process knowledge graph, three models were established to solve the co-occurrence probability of candidate entity groups, the overall semantic similarity of candidate entity groups, and the conditional probability between entity mentions and candidate entities, respectively (1), 2), and 3 in this section).

[0100] 1) Overall correlation measurement model for candidate entity groups based on strongly correlated sequences

[0101] This section uses the overall relevance of candidate entity groups to estimate their co-occurrence probability. The relevance of candidate entity groups is based on the relevance of entity pairs. Therefore, to accurately measure the relevance of entity pairs, this section establishes an entity relevance measurement model based on three 2-hop paths. Furthermore, to emphasize the contribution of highly correlated entity pairs to the overall relevance of the entire candidate entity group, this section establishes an overall relevance measurement model for candidate entity groups based on strongly correlated sequences to capture the overall relevance.

[0102] ① Entity relevance measurement model based on three 2-hop paths

[0103] Studies have shown that 1-hop and 2-hop paths between entities can effectively capture the correlation information between entities. However, longer paths between entities introduce more noise, which reduces the correlation information between entities. To improve the information utilization rate in the knowledge graph of UAV engine disassembly process, this section divides the 2-hop paths in the process knowledge graph into three types: the first type points to a common process entity; the second type is pointed to by a common process entity; and the last type is a directed 2-hop path formed between two process entities.

[0104] like Figure 2 As shown, there are three 2-hop paths between the process entities "Right Side Cover [Cylinder Head]" and "Cylinder Head". In path A, "Right Side Cover [Cylinder Head] → Bolt [Left Cylinder Cover] ← Cylinder Head" is the first type of 2-hop path, where the relationship is that of a preceding part; in path B, "Right Side Cover [Cylinder Head] ← Hand Use → Cylinder Head" is the second type of 2-hop path, where the relationship is that a tool is used; and in path C, "Cylinder Head → Seal [Right Cylinder Head Cover] → Right Side Cover [Cylinder Head]" is the third type of 2-hop path, where the relationship is that of a preceding part. All three paths contribute to the relevance of "Right Side Cover [Cylinder Head]" and "Cylinder Head". Therefore, it is necessary to consider the contribution of all three 2-hop paths to the entity's relevance simultaneously.

[0105] Generally, the shorter the path between process entities, the greater the correlation between them; the longer the path, the smaller the correlation. The more 2-hop paths between process entities, the greater their correlation. Therefore, this section proposes using three types of 2-hop paths and one-hop paths (i.e., two candidate entities directly connected) to more accurately measure the correlation between process entities. Considering the different numbers of each type of 2-hop path and the varying importance of each 2-hop path to the correlation of process entities, this section establishes an entity correlation measurement model based on three types of 2-hop paths, as shown in formula (3).

[0106]

[0107] In the formula R(e i ,e j — Process entities ei and e j The correlation measurement function;

[0108] (e i ,r,e j — In the knowledge graph, the process entity e i and e j There is a relationship between them;

[0109] R2(e i ,e j— A correlation measurement function based on 2-hop paths;

[0110] ——Process Entity e i to e j The path length, in this section

[0111] —respectively, process entity e i The number of process entities pointed to and e j The number of process entities it points to;

[0112] — Representing process entity e respectively i With e j The number of first-type 2-hop paths and the number of second-type 2-hop paths;

[0113] ——Process Entity e i to e j The third type of 2-hop path quantity;

[0114] β—a hyperparameter that measures the importance of a directed 2-hop path to other 2-hop paths;

[0115] ② Overall correlation measurement model for candidate entity groups based on strongly correlated sequences

[0116] Knowledge graphs are constructed from a large amount of text or prior knowledge in the human mind. Therefore, the candidate entity group corresponding to an entity mention group can also be regarded as a whole. When calculating candidate entities e1, e2, ..., e... q The co-occurrence probability P(e1,e2,…,e q When (e1,e2,…,e) q The stronger the overall relevance of candidate entity groups e1, e2, ..., e q The higher the co-occurrence probability, the more challenging it is to effectively measure the overall relevance of candidate entity groups. To mitigate the impact of unrelated or weakly related entity pairs on the overall relevance and to emphasize the contribution of strongly related entity pairs, the concepts of relevance graphs and strongly related sequences are defined.

[0117] Definition 5: Relevance Graph: A candidate entity group (e1, e2, ..., e) for a given entity mention group. q ) and the correlation R(e) between any two candidate entities i ,e j A q-order correlation graph is an undirected complete graph formed by connecting all entities in a candidate entity group as nodes and the correlation between any two entities as edges between entities.

[0118] Definition 6. Correlation Sequence: If a correlation graph is a q-order correlation graph, then a subgraph that satisfies three conditions can be called a correlation sequence: (1) the subgraph has q-1 edges; (2) the subgraph is a connected graph; and (3) the subgraph is a generating subgraph of the q-order correlation graph. Among all correlation sequences of a correlation graph, the correlation sequence with the highest cumulative correlation is called a strong correlation sequence.

[0119] Figure 3 This represents a 4th-order correlation graph and its two correlation sequences. The red correlation sequence has the highest correlation among all sequences and is called a strongly correlated sequence. By accumulating the correlations on the strongly correlated sequence, we obtain a correlation of 2.4, i.e., 0.7 + 0.8 + 0.9 = 2.4. According to the definition of a correlation sequence, it is not necessarily a path; it could also be a tree. Therefore, a strongly correlated sequence may ultimately be a tree. For example, in... Figure 3 The correlation sequences, represented by the solid blue line, form a tree. Strongly correlated sequences can be solved using the maximum weighted spanning tree algorithm.

[0120] Based on strongly correlated sequences, this section constructs an overall correlation measurement model for candidate entity groups based on strongly correlated sequences, as shown in formula (4).

[0121]

[0122] In the formula, Tree1(e1,e2,…,e q — Entities e1, e2, ..., e are obtained based on the relevance of entity pairs. q Strongly correlated sequences;

[0123] (e i ,e j — An edge in a strongly correlated sequence;

[0124] R(e i ,e j )——Entity e i ,e j The correlation.

[0125] 2) Overall semantic similarity model of candidate entity groups

[0126] How to effectively obtain the semantic information of process entities is a difficult problem in measuring the overall semantic similarity of candidate entity groups. Knowledge graph embedding methods can effectively learn the contextual information of entities, thereby extracting the semantic features of entities and relationships, and converting the text description of entities into semantic vectors. For example, ROTATE, ERDERP model, etc. Based on the semantic vector of process entities, the semantic similarity between any two process entities can be quantified by cosine similarity or Euclidean distance, as shown in formula (5). For example, after ERDERP model, the semantic vector of process entity "starting shaft" is [-0.17818,0.13530,…,-0.00888,0.23183], and the semantic vector of process entity "crankshaft connecting rod piston combination" is [-0.08544,0.12662,…,-0.02125,0.03360], and the cosine similarity is 0.831, that is, the semantic similarity between the two entities is 0.831. As shown in formula (4).

[0127]

[0128] In the formula ——Process Entity e i semantic vector;

[0129] —Calculation of process entity and process entity Semantic similarity functions, such as: cosine similarity function, reciprocal of Euclidean distance;

[0130] Similarly, to further highlight the strong logical connections between process knowledge, a strongly correlated sequence is used to calculate the overall semantic similarity of candidate entity groups, as shown in the formula.

[0131]

[0132] In the formula, Tree2(e1,e2,…,e q — Entities e1, e2, ..., e are obtained based on the semantic similarity of entity pairs. q Strongly correlated sequences;

[0133] 3) Surface similarity model

[0134] Calculate P(m) using word frequency i |e i This is an effective method. However, we cannot obtain all candidates e. i and entity mention m i Related web documents. On the one hand, in the online world, entity e i and entity mention m iInformation is constantly changing; on the other hand, due to the nature of complex equipment disassembly processes, some entities and the information they mention are difficult to find on the network. Therefore, it is not possible to directly use word frequency-based methods to calculate P(m). i |e i In fact, entity mentions entered by disassembly engineers generally have a high degree of surface similarity to correct candidate entities. Therefore, this section assumes that entity mention m i and candidate entity e i The greater the surface similarity, the more P(m) i |e i The larger the value, the greater the value. Specifically, edit distance is used to calculate entity mention m. i and candidate entity e i The surface similarity is used to estimate P(m). i |e i As shown in formula (6).

[0135]

[0136] In the formula |m i |、|e i |——m refers to the entity respectively i and candidate entity e i The number of characters;

[0137] lev(m i ,e i — Entity mention m i and candidate entity e i Edit distance;

[0138] Lv(m i ,e i — Entity mention m i and candidate entity e i Surface similarity.

[0139] Formula (6) shows that as long as there exists i such that S(m) i ,e i If ) = 0, then P(m1,m2,…,m q |e1,e2,…,e q The value of ) will equal 0, resulting in the entity mention group (m1, m2, ..., m) being reduced to 0. q Link to candidate entity group (e1, e2, ..., e qThe probability of an entity mention being linked to by another entity mention is 0. If multiple candidate entity groups have a probability of 0, it will be difficult to distinguish the link probabilities of entity mentions linking to different candidate entity groups. To avoid this situation, a hyperparameter γ greater than 0 is added when calculating the surface similarity between entity mentions and candidate entities, as shown in formula (7).

[0140]

[0141] (1) Experimental verification

[0142] 1) Experimental data

[0143] To verify the effectiveness of the fine-grained collective entity linking method for disassembly process knowledge graphs proposed in this chapter, the actual disassembly process dataset of the JH70 engine (EDPK) was used as the entity linking sample, and the constructed JH70 engine disassembly process knowledge graph was used as the knowledge base for entity linking. Entity linking experiments were conducted. To verify the generalization ability of the proposed collective entity linking method, the publicly available datasets AIDA CoNLL-YAGO, ACE2004, and AQUAINT were used as the entity linking data, and a portion of the data in YAGO-Core was used as the knowledge base for entity linking. Entity linking experiments were conducted on YAGO-Core. Dataset details are shown in Table 1.

[0144] Table 1 Experimental Data

[0145]

[0146] 2) Model Evaluation

[0147] In the experiments, precision, recall, and F1 score were used to evaluate the entity linking performance of the entity linking method on various datasets.

[0148]

[0149] In the formula, `correct_link` represents the number of entity mentions in the correct link.

[0150] processed_mentions — The number of entity mentions that have candidate entities.

[0151] total_mentions — The number of times the linked entity is mentioned in the knowledge graph.

[0152] 3) Comparison methods and experimental setup

[0153] In the experiment, the SRSCL method was compared with five other entity linking methods on the aforementioned dataset. These five methods included the classic Entity Popularity Model (POP), and the latest collective entity linking methods: CFEL, EMDD, TTHP, and CFELS. The Entity Popularity Model calculates the popularity of each entity using its degree information and ranks candidate entities based on popularity to obtain the most likely linked candidate entities. CFEL calculates entity relevance using directed 2-hop paths between entities and uses this relevance to characterize the co-occurrence probability of entities, finally using a probabilistic model to complete the entity linking. Within the CFEL framework, TTHP uses three 2-hop paths to measure entity relevance; EMDD uses semantic vectors to measure the relevance of entity pairs; and CFELS introduces strongly correlated sequences to achieve collective entity linking.

[0154] Parameter tuning experiments show that, for entity linking on the engine disassembly process knowledge graph, the SRSCL method achieves the best results when the sliding window length is set to 20, the number of candidate entities is set to top-10, β is set to 0.8, and γ is set to 0.5. Repeating the above experiments, it was found that for entity linking on YAGO-Core, the SRSCL method achieves the best results when the sliding window length is set to 15, the number of candidate entities is set to top-10, β is set to 0.3, and γ is set to 0.4.

[0155] 4) Comparative Experiment

[0156] To verify the effectiveness of the proposed SRSCL method in engine disassembly process knowledge retrieval, a collective entity linking experiment was conducted using the JH70 engine disassembly process knowledge graph as the knowledge base and EDPK as the dataset. The results are shown in Table 2. Simultaneously, EMDD, TTHP, CFELS, and SRSCL were compared with CFEL. The entity linking precision, recall, and F1 score of the five models are shown in Table 2. Figure 4 As shown.

[0157] Table 2 Comparative experimental results of entity linking methods on EDPK

[0158]

[0159] Table 4 shows that the SRSCL model proposed in this invention significantly outperforms other models on all evaluation metrics. For example... Figure 4As shown, TTHP outperforms CFEL in precision, recall, and F1 score. By comparing TTHP with CFEL, the results verify that the three 2-hop paths can more comprehensively represent the features of process entities, thus significantly improving the performance of entity linking. EMDD outperforms CFEL in all evaluation metrics. Experimental results show that semantic vectors can effectively extract deep features of process entities compared to directed 2-hop paths, thereby improving the performance of entity linking. CFELS outperforms CFEL in precision, recall, and F1 score, verifying that strongly correlated sequences can effectively mine the correlation information between process entities, thus significantly improving the precision of entity linking. The SRSCL model proposed in this invention integrates 2-hop paths, semantic vectors, and strongly correlated sequences. The method of this invention can effectively support knowledge retrieval on the knowledge graph of UAV engine disassembly process.

[0160] 5) SRSCL generalization verification experiment

[0161] To verify the generalization ability of the proposed SRSCL, entity linking experiments were conducted on the public datasets AIDA CoNLL-YAGO, ACE2004, and AQUAINT, with a portion of the data from YAGO-Core used as the knowledge base for entity linking. Precision, recall, and F1 score were used to evaluate the entity linking performance of the methods. The entity linking results of the six methods on the three public datasets are shown in Table 3.

[0162] Table 3 shows the comparative experimental results of the entity linking method on three public datasets.

[0163]

[0164] Table 3 shows that, compared to the other five models, the SRSCL method proposed in this chapter achieves the best precision, recall, and F1 score on all three datasets. Experimental results demonstrate that the proposed SRSCL can effectively complete entity linking tasks on public datasets. This verifies that the proposed SRSCL has strong generalization ability and can provide a reference for entity linking in other and general domains. Detailed analysis is shown below.

[0165] In terms of accuracy on AIDA CoNLL-YAGO, SRSCL improved by 0.3%-7.24% compared to the other five comparison methods, with an average improvement of 2.77%. Simultaneously, SRSCL significantly outperformed the comparison methods in recall and F1 score. On AQUAINT, POP achieved the worst results across all evaluation metrics, with CFEL achieving the second worst. SRSCL, EMDD, TTHP, and CFELS significantly outperformed CFEL, with SRSCL achieving the best results. Furthermore, SRSCL significantly outperformed POP and CFEL; for example, SRSCL's F1 score was 14.01% and 3.37% higher than POP and CFEL, respectively. Experimental results on ACE2004 showed that due to the sparsity of the ACE2004 data, multiple models exhibited similar entity linking performance on ACE2004. The POP model had the worst entity linking performance, CFEL, EMDD, and CFELS all achieved the second best results, and TTHP and SRSCL both achieved the best results. Experimental results verify that the SRSCL proposed in this chapter can effectively link entities in public datasets, and that SRSCL has strong generalization ability.

[0166] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. 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 semantic retrieval method for a knowledge graph of unmanned aerial vehicle (UAV) engine disassembly process, wherein the method applies to the text description of the actual engine disassembly process, and uniquely locates the text description based on a given knowledge graph, characterized in that... Select the entity mention set in the text description. Where q represents the number of entity mentions, and entity mentions are generated using a candidate entity generation method. The generated candidate entity set is ,in, Indicates entity mention The One candidate entity, Indicates entity mention The number of candidate entities, Candidate entity generation: finding mentions of each entity from a given knowledge graph. Possible candidate entity set , Noise removal is achieved through two pruning processes. First, candidate entities are pruned based on their mention categories. Then, a second pruning is performed based on the surface similarity between the candidate entities and their mentions. Objective function construction: Construct a collective entity linking objective function for the knowledge graph of UAV engine disassembly process; To solve the objective function, three models were established to solve the co-occurrence probability of candidate entity groups, the overall semantic similarity of candidate entity groups, and the conditional probability between entity mentions and candidate entities, respectively. Wherein, the candidate entity group is an element in the Cartesian product of all candidate entity sets; the co-occurrence probability of the candidate entity group is the probability that all candidate entities in the candidate entity group appear together in the same text segment; the overall semantic similarity of the candidate entity group is the overall semantic similarity among all candidate entities in the candidate entity group; and the conditional probability between an entity mention and a candidate entity is the probability that an entity mention will appear given a candidate entity.

2. The semantic retrieval method for a knowledge graph of UAV engine disassembly process according to claim 1, characterized in that, In the construction of the objective function, the specific formula is as follows: in, Indicates that given a section of process search text, One entity was mentioned; Indicates entity mention candidate entity set The first in One candidate entity; Represents the Cartesian product of the set of candidate entities mentioned by all entities; Indicates entity mention group A candidate entity group; Indicates entity mention group Link to candidate entity group The probability of; Representing entities The probability of appearing in the same text segment; This means that in a text, given candidate entities... When the entity mentions The probability of occurrence.

3. The semantic retrieval method for a knowledge graph of UAV engine disassembly process according to claim 2, characterized in that, Accurately estimate based on the overall semantic similarity of entity mention groups and candidate entity groups The specific formula is as follows: in, Indicates the number of entity references in a text; Indicates candidate entity group The overall semantic similarity function; Represents a given candidate entity Below, entity mentions The probability of occurrence The entity link objective function is expressed as follows: 。 4. The semantic retrieval method for a knowledge graph of UAV engine disassembly process according to claim 1, characterized in that, The co-occurrence probability of candidate entity groups is estimated by using the overall relevance of candidate entity groups, and an entity relevance measurement model based on three 2-hop paths is established. in, Represents the process entity and The correlation measurement function; In the knowledge graph, the process entity and There is a relationship between them; This represents a correlation measurement function based on 2-hop paths; Represents the process entity arrive The path length, in this section ; , They are respectively the process entities The number of process entities pointed to and The number of process entities it points to; , Representing the process entity and The number of first-type 2-hop paths and the number of second-type 2-hop paths; Represents the process entity arrive The third type of 2-hop path quantity; This represents a hyperparameter that measures the importance of a directed 2-hop path to other 2-hop paths; The first type of 2-hop path refers to the path where two process entities point to a common process entity. The second type of 2-hop path refers to the path where two process entities are pointed to by a common process entity. The third type of 2-hop path refers to the directed 2-hop path formed between two process entities.

5. A semantic retrieval method for a knowledge graph of UAV engine disassembly process according to claim 4, characterized in that, Construct an overall correlation measurement model for candidate entity groups based on strongly correlated sequences. ； in, Represents entities obtained based on the relevance of entity pairs. The strongly correlated sequence is a generating subgraph in the correlation graph that satisfies the following conditions: 1) the subgraph has q-1 edges; 2) the subgraph is a connected graph; 3) the subgraph is a generating subgraph of a q-order correlation graph, and has the highest cumulative correlation among all correlation sequences; Represents an edge in a strongly correlated sequence; Representing entities The correlation.

6. A semantic retrieval method for a knowledge graph of UAV engine disassembly process according to claim 4, characterized in that, The solution for the overall semantic similarity of candidate entity groups includes quantifying the semantic similarity between any two process entities using cosine similarity or Euclidean distance based on the semantic vectors of the process entities. ， For the process entity semantic vector; For the calculation of process entities and process entity The semantic similarity function.

7. A semantic retrieval method for a knowledge graph of UAV engine disassembly process according to claim 6, characterized in that, Strongly correlated sequences are used to calculate the overall semantic similarity of candidate entity groups, highlighting the strong logical connections between process knowledge. ， in, Represents entities obtained based on the semantic similarity of entity pairs. The strongly correlated sequence is a generated subgraph in the correlation graph that satisfies the following conditions: 1) the subgraph has q-1 edges; 2) the subgraph is a connected graph; 3) the subgraph is a generated subgraph of the q-order correlation graph and has the largest cumulative semantic similarity among all the correlation sequences.

8. A semantic retrieval method for a knowledge graph of UAV engine disassembly process according to claim 4, characterized in that, Entity mentions are calculated using edit distance. and candidate entities The surface similarity is used to estimate the conditional probability between entity mentions and candidate entities. ， ， in, , Entity mentions respectively and candidate entities The number of characters; Indicates entity mention and candidate entities Edit distance; Indicates entity mention and candidate entities Surface similarity.

9. A semantic retrieval method for a knowledge graph of UAV engine disassembly process according to claim 8, characterized in that, When calculating the surface similarity between entity mentions and candidate entities, add a hyperparameter greater than 0. , 。

Images