A knowledge graph completion method and system based on multi-view link prediction

Abstract

This specification provides a knowledge graph completion method and system based on multi-view link prediction. The method includes: establishing an existing knowledge concept automatic converter; establishing and training a dual encoder; obtaining the triple to be predicted; the existing knowledge concept automatic converter determining the candidate tail entities and candidate tail entity text descriptions corresponding to the triple to be predicted based on the head entities and relations included in the triple to be predicted; the dual encoder determining the score of the candidate tail entities based on the candidate tail entities and candidate tail entity text descriptions; and determining the tail entities included in the triple to be predicted based on the scores of the candidate tail entities. This method has the advantage of improving the accuracy of knowledge graph completion.

Description

Technical Field

[0001] This specification relates to the field of knowledge graphs, and in particular to a knowledge graph completion method and system based on multi-view link prediction. Background Technology

[0002] The entire lifecycle of avionics equipment includes design models, manufacturing models, and operation and maintenance models. These models exhibit significant differences in data structure, content, and standards, and their data is largely independent. Constructing a distributed integrated design, manufacturing, and operation and maintenance platform can effectively integrate diverse and heterogeneous data, addressing data collection and governance across different equipment, production lines, instruments, and information systems. This enables unified construction and data fusion of heterogeneous models from multiple enterprises and products, facilitating parallel processing of complex business datasets across multiple enterprise supply chains and disciplines. It also allows for the cross-business and cross-enterprise integration and inversion of digital models. Among these, multi-business model data fusion is particularly crucial in the construction of the distributed integrated design and manufacturing platform. It integrates collected design, manufacturing, and operation and maintenance data to achieve a unified data source. This unified data source can then be used for data analysis applications. During data fusion, a knowledge graph needs to be constructed to support data fusion and storage. Currently, knowledge graphs are primarily constructed manually or semi-automatically, making it difficult to uncover some implicit relationships. Therefore, supplementing the knowledge graph is essential to ensure its completeness.

[0003] Therefore, there is a need to provide a knowledge graph completion method and system based on multi-view link prediction to improve the accuracy of knowledge graph completion. Summary of the Invention

[0004] One embodiment of this specification provides a knowledge graph completion method based on multi-view link prediction. The method includes: establishing an existing knowledge concept automatic converter; establishing and training a dual encoder; obtaining a triple to be predicted; the existing knowledge concept automatic converter determining candidate tail entities and candidate tail entity text descriptions corresponding to the triple to be predicted based on the head entities and relations included in the triple to be predicted; the dual encoder determining the score of the candidate tail entities based on the candidate tail entities and candidate tail entity text descriptions; and determining the tail entities included in the triple to be predicted based on the score of the candidate tail entities.

[0005] In some embodiments, the establishment of an existing knowledge concept automatic converter includes: establishing an entity-to-existing knowledge concept mapping dictionary and an entity-to-entity text description mapping dictionary; using the entity-to-existing knowledge concept mapping dictionary and the entity-to-entity text description mapping dictionary, converting triples in the triple dataset into existing knowledge concept triples, and establishing an existing knowledge concept triple dataset corresponding to the triple dataset.

[0006] In some embodiments, establishing and training the dual encoder includes: establishing an initial dual encoder, wherein the initial dual encoder includes a first BERT model and a second BERT model; training the initial dual encoder using positive and negative samples to generate the dual encoder.

[0007] In some embodiments, obtaining the negative samples includes: generating hard negative samples and soft negative samples; and filtering the soft negative samples based on existing knowledge.

[0008] In some embodiments, the existing knowledge concept automatic converter determines the candidate tail entity and candidate tail entity text description corresponding to the triplet to be predicted based on the head entity and relations included in the triplet to be predicted, including: determining the candidate existing knowledge concept set corresponding to the tail entity included in the triplet to be predicted based on the existing knowledge concept set corresponding to the head entity included in the triplet to be predicted and the relations included in the triplet to be predicted; the entity and entity text description mapping dictionary determines the candidate tail entity and candidate tail entity text description corresponding to the triplet to be predicted based on the candidate existing knowledge concept set corresponding to the tail entity.

[0009] In some embodiments, the dual encoder determines the score of the candidate tail entity based on the candidate tail entity and the candidate tail entity text description, including: the first BERT model determines the head entity relation-aware embedding based on the head entity text description corresponding to the head entity in the triplet to be predicted and the relation text description corresponding to the relation; the second BERT model determines the candidate tail entity embedding based on the candidate tail entity text description; and the score of the candidate tail entity is determined based on the similarity between the head entity relation-aware embedding and the candidate tail entity embedding.

[0010] One embodiment of this specification provides a knowledge graph completion system based on multi-view link prediction. The system includes: a converter establishment module for establishing an automatic converter of existing knowledge concepts; an encoder establishment module for establishing and training a dual encoder; a triple acquisition module for acquiring triples to be predicted; and a triple prediction module for using the automatic converter of existing knowledge concepts to determine candidate tail entities and candidate tail entity text descriptions corresponding to the triples to be predicted based on the head entities and relationships included in the triples to be predicted. The triple prediction module is further used to use the dual encoder to determine the score of the candidate tail entities based on the candidate tail entities and candidate tail entity text descriptions. The triple prediction module is also used to determine the tail entities included in the triples to be predicted based on the scores of the candidate tail entities.

[0011] In some embodiments, the converter establishment module is further configured to: establish an entity-to-existing knowledge concept mapping dictionary and an entity-to-entity text description mapping dictionary; convert the triples in the triple dataset into existing knowledge concept triples through the entity-to-existing knowledge concept mapping dictionary and the entity-to-entity text description mapping dictionary, and establish an existing knowledge concept triple dataset corresponding to the triple dataset.

[0012] In some embodiments, the encoder building module is further configured to: build an initial dual encoder, wherein the initial dual encoder includes a first BERT model and a second BERT model; train the initial dual encoder using positive and negative samples to generate the dual encoder.

[0013] In some embodiments, the triple prediction module is further configured to: determine the head entity relation-aware embedding based on the head entity text description corresponding to the head entity in the triple to be predicted and the relation text description corresponding to the relation; determine the candidate tail entity embedding based on the candidate tail entity text description; and determine the score of the candidate tail entity based on the similarity between the head entity relation-aware embedding and the candidate tail entity embedding. Attached Figure Description

[0014] This specification will be further described by way of exemplary embodiments, which will be described in detail with reference to the accompanying drawings. These embodiments are not limiting; in these embodiments, the same reference numerals denote the same structures, wherein:

[0015] Figure 1 This is a block diagram of an exemplary knowledge graph completion system based on multi-view link prediction, as shown in some embodiments of this specification;

[0016] Figure 2 This is a flowchart illustrating an exemplary knowledge graph completion method based on multi-view link prediction according to some embodiments of this specification;

[0017] Figure 3 This is a flowchart illustrating an exemplary method for obtaining negative samples according to some embodiments of this specification;

[0018] Figure 4 This is a flowchart illustrating the exemplary determination of the score of a candidate tail entity according to some embodiments of this specification. Detailed Implementation

[0019] To more clearly illustrate the technical solutions of the embodiments in this specification, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are merely some examples or embodiments of this specification. For those skilled in the art, these drawings can be applied to other similar scenarios without creative effort. Unless obvious from the context or otherwise specified, the same reference numerals in the drawings represent the same structures or operations.

[0020] It should be understood that the terms “system,” “device,” “unit,” and / or “module” used herein are one way to distinguish different components, elements, parts, sections, or assemblies at different levels. However, if other terms can achieve the same purpose, they may be replaced by other expressions.

[0021] As indicated in this specification and claims, unless the context clearly indicates otherwise, the words "a," "an," "an," and / or "the" do not specifically refer to the singular and may also include the plural. Generally speaking, the terms "comprising" and "including" only indicate the inclusion of expressly identified steps and elements, which do not constitute an exclusive list, and the method or apparatus may also include other steps or elements.

[0022] Flowcharts are used in this specification to illustrate the operations performed by the system according to embodiments of this specification. It should be understood that the preceding or following operations are not necessarily performed in exact order. Instead, the steps can be processed in reverse order or simultaneously. Furthermore, other operations can be added to these processes, or one or more steps can be removed from them.

[0023] Figure 1 This is a module diagram of an exemplary knowledge graph completion system based on multi-view link prediction, as shown in some embodiments of this specification. Figure 1 As shown, a knowledge graph completion system based on multi-view link prediction can include a converter establishment module, an encoder establishment module, a triple acquisition module, and a triple prediction module.

[0024] The converter building module can be used to build automatic converters of existing knowledge concepts.

[0025] In some embodiments, the converter establishment module can also be used to: establish an entity-to-existing knowledge concept mapping dictionary and an entity-to-entity text description mapping dictionary; convert the triples in the triple dataset into existing knowledge concept triples through the entity-to-existing knowledge concept mapping dictionary and the entity-to-entity text description mapping dictionary, and establish an existing knowledge concept triple dataset corresponding to the triple dataset.

[0026] The encoder building module can be used to build and train dual encoders.

[0027] In some embodiments, the encoder building module can also be used to: build an initial dual encoder, wherein the initial dual encoder includes a first BERT model and a second BERT model; train the initial dual encoder using positive and negative samples to generate the dual encoder. In some embodiments, the encoder building module can also be used to: generate hard negative samples and soft negative samples; and filter the soft negative samples based on existing knowledge.

[0028] The triplet acquisition module can be used to obtain triples to be predicted.

[0029] The triplet prediction module can be used to determine the candidate tail entities and their textual descriptions based on the head entities and relationships included in the triplet to be predicted, using existing knowledge-based automatic concept converters. The triplet prediction module can also be used to determine the scores of the candidate tail entities based on their textual descriptions using a dual encoder. Furthermore, the triplet prediction module can be used to determine the tail entities included in the triplet to be predicted based on their scores.

[0030] In some embodiments, the triple prediction module can also be used to: determine the candidate set of existing knowledge concepts corresponding to the head entities included in the triple to be predicted based on the set of existing knowledge concepts corresponding to the head entities included in the triple to be predicted and the relationship included in the triple to be predicted; and use an entity-entity text description mapping dictionary to determine the candidate tail entities and candidate tail entity text descriptions corresponding to the triple to be predicted based on the candidate set of existing knowledge concepts corresponding to the tail entities.

[0031] In some embodiments, the triple prediction module can also be used for: a first BERT model determining the head entity relation-aware embedding based on the head entity text description corresponding to the head entity and the relation text description corresponding to the relation in the triple to be predicted; a second BERT model determining the candidate tail entity embedding based on the candidate tail entity text description; and determining the score of the candidate tail entity based on the similarity between the head entity relation-aware embedding and the candidate tail entity embedding.

[0032] For more detailed descriptions of the converter creation module, encoder creation module, triplet acquisition module, and triplet prediction module, please refer to [link to relevant documentation]. Figure 2 The details and related descriptions will not be repeated here.

[0033] It should be noted that the above description of the knowledge graph completion system and its modules based on multi-view link prediction is for ease of description only and should not be construed as limiting this specification to the scope of the embodiments described. It is understood that those skilled in the art, after understanding the principles of this system, may arbitrarily combine the various modules or construct subsystems connected to other modules without departing from these principles. In some embodiments, Figure 1 The converter establishment module, encoder establishment module, triplet acquisition module, and triplet prediction module disclosed herein can be different modules within a single system, or a single module can implement the functions of two or more of the aforementioned modules. For example, the modules can share a single storage module, or each module can have its own dedicated storage module. Such variations are all within the scope of protection of this specification.

[0034] Figure 2 This is a flowchart illustrating an exemplary knowledge graph completion method based on multi-view link prediction, according to some embodiments of this specification. In some embodiments, the knowledge graph completion method based on multi-view link prediction can be executed by a knowledge graph completion system based on multi-view link prediction. For example, the knowledge graph completion method based on multi-view link prediction can be stored in a storage device in the form of a program or instructions, and the knowledge graph completion method based on multi-view link prediction can be implemented when the knowledge graph completion system based on multi-view link prediction executes the instructions. The operational schematic diagrams of the knowledge graph completion method based on multi-view link prediction presented below are illustrative. In some embodiments, the process can be accomplished using one or more additional operations not described and / or one or more operations not discussed. Additionally, Figure 2 The order of operations shown in the diagram and described below for the knowledge graph completion method based on multi-view link prediction is not restrictive. Figure 2 As shown, the knowledge graph completion method based on multi-view link prediction can include the following steps.

[0035] Step 210: Establish an automatic converter for existing knowledge concepts. In some embodiments, step 210 may be performed by a converter establishment module.

[0036] In some embodiments, existing knowledge can be knowledge that already exists, such as common sense.

[0037] In some embodiments, the converter establishment module establishes an automatic converter of existing knowledge concepts, which may include:

[0038] Establish a dictionary mapping entities to existing knowledge concepts and a dictionary mapping entities to textual descriptions.

[0039] By using the entity-to-existing knowledge concept mapping dictionary and the entity-to-entity text description mapping dictionary, the triples in the triple dataset are transformed into existing knowledge concept triples, and an existing knowledge concept triple dataset corresponding to the triple dataset is established.

[0040] For example, the triple (Xiaomi, founder, Lei Jun) can be transformed into the corresponding existing knowledge concept triple dataset (company, founder, person).

[0041] Step 220: Establish and train the dual encoder. In some embodiments, step 220 may be performed by the encoder establishment module.

[0042] Dual encoders are used to vectorize entity descriptions and relation descriptions.

[0043] In some embodiments, the encoder building module builds and trains a dual encoder, which may include:

[0044] An initial dual encoder is established, which includes a first BERT (Bidirectional Encoder Representation from Transformers) model and a second BERT model;

[0045] The initial dual encoder is trained using positive and negative samples, and then the dual encoder is generated.

[0046] In some embodiments, the encoder building module for acquiring negative samples may include:

[0047] Generate hard negative samples and soft negative samples;

[0048] Based on existing knowledge, soft negative samples are screened.

[0049] Understandably, the ability to construct a large number of high-quality negative samples is crucial for model training. A major reason why text-based methods currently lag behind embedding-based methods is that the latter do not involve the high computational costs of text editors and can train on more negative samples simultaneously. Using dual encoders can address the high computational cost issue, making it especially important to obtain as many high-quality negative samples as possible. To obtain high-quality negative samples, existing knowledge concepts (e.g., common sense concepts) are incorporated into the construction of the negative samples. The constructed negative samples are divided into soft negative samples and hard negative samples.

[0050] Hard negative samples are always correct negative samples, and they mainly come from two parts: Figure 3 This is a flowchart illustrating exemplary methods for obtaining negative samples according to some embodiments of this specification, such as... Figure 3As shown, this includes samples within the same training batch and samples from the first 1-2 training batches. The relationships between entities can be categorized into four types: one-to-one (…). One-to-many Many to one () Many-to-many In this context, 1 represents the condition that when one entity and relation of a triple are known, the other entity in the triple is uniquely determined. The opposite meaning is 'n', representing the condition that when one entity and relation of a triple are known, the other entity in the triple is not unique. We construct negative samples using samples from the same training batch and the first 1-2 training batches. First, we extract the relation as... , , For a triplet sample, the uniquely identified entity in it is replaced with any entity from the same batch or the previous 1-2 batches that is different from the replaced entity. The resulting negative sample is always a correct negative sample.

[0051] Soft negative samples replace non-unique entities (i.e., the entity corresponding to n) with entities from the same batch or the previous 1-2 batches of samples. Obviously, since n is not uniquely determined, soft negative samples may actually be incorrect negative samples (i.e., positive samples). In contrast to hard negative samples, which are extremely beneficial for model training, incorrect negative samples are extremely detrimental to model training, causing the model to be trained in the wrong direction. Therefore, reducing incorrect negative samples is crucial. This is where existing knowledge concepts (e.g., common sense concepts) come in. Figure 3 As shown, samples that conform to existing knowledge concepts (e.g., common sense concepts) in the soft negative samples are discarded. In theory, erroneous negative samples must satisfy existing knowledge concepts (e.g., common sense concepts). Therefore, after filtering through the set of triples of existing knowledge concepts (e.g., common sense concepts), the remaining soft negative samples are all correct negative samples.

[0052] Both the first and second BERT models are initialized using the same pre-trained BERT language model, but they do not share parameters. The two pre-trained BERT language models are trained using hard negative samples, soft negative samples obtained through commonsense-based negative sampling, and positive samples from the triplet dataset. The two pre-trained BERT language models are then fine-tuned through positive-negative sample contrast learning, thus generating the first and second BERT models.

[0053] Step 230: Obtain the triples to be predicted. In some embodiments, step 230 may be performed by the triple acquisition module.

[0054] Step 240: The existing knowledge concept automatic converter determines the candidate tail entities and candidate tail entity text descriptions corresponding to the triples to be predicted based on the head entities and relations included in the triples to be predicted. In some embodiments, step 240 can be performed by the triple prediction module.

[0055] Understandably, both embedding-based and text-based knowledge graph completion techniques ultimately require link prediction. Addressing the link uncertainty inherent in this prediction stage is crucial, as it significantly impacts the accuracy of knowledge graph completion. This paper proposes a multi-perspective link prediction process, progressing from coarse to fine, by introducing existing knowledge concepts (e.g., common sense concepts).

[0056] In some embodiments, coarse-view link prediction may include: determining a set of candidate existing knowledge concepts corresponding to the tail entities included in the triplet to be predicted, based on the set of existing knowledge concepts corresponding to the head entities and the relations included in the triplet to be predicted; and determining the candidate tail entities and candidate tail entity text descriptions corresponding to the triplet to be predicted based on the candidate set of existing knowledge concepts corresponding to the tail entities using an entity-entity text description mapping dictionary. Specifically, when predicting the tail entity of the triplet to be predicted (head entity, relation, tail entity), all entities in the existing knowledge concept triplet dataset are candidate entities. The set of candidate existing knowledge concepts corresponding to the tail entities is first obtained through the set of existing knowledge concepts corresponding to the head entities and the relations, thus performing a coarse filtering of the candidate entities. The set of candidate existing knowledge concepts corresponding to the tail entities is shown in the following formula:

[0057]

[0058] in This represents the set of candidate existing knowledge concepts corresponding to the tail entity. This represents the j-th existing knowledge concept corresponding to the head entity. Common sense The tail entity in the text corresponds to the j-th existing knowledge concept. This represents common sense. By leveraging existing knowledge concepts of entities and the mapping relationship between entities and their descriptions, coarsely filtered candidate tail entities and their textual descriptions are obtained.

[0059] Step 250: The dual encoder determines the score of the candidate tail entity based on the candidate tail entity and its text description. In some embodiments, step 250 can be performed by the triplet prediction module.

[0060] Figure 4 This is a flowchart illustrating exemplary methods for determining the score of candidate tail entities according to some embodiments of this specification, such as... Figure 4As shown, in some embodiments, fine-grained link prediction may include: a first BERT model determining the head entity relation-aware embedding based on the head entity text description corresponding to the head entity and the relation text description corresponding to the relation in the triple to be predicted; a second BERT model determining the candidate tail entity embedding based on the candidate tail entity text description; and determining the score of the candidate tail entity based on the similarity between the head entity relation-aware embedding and the candidate tail entity embedding. Specifically, to achieve the final link prediction, the triple to be predicted and the candidate entities are processed as follows:

[0061] The first BERT model calculates the relation-aware embedding of the head entity in the triplet to be predicted. The input of the first BERT model can include the text descriptions corresponding to the head entities in the triplet to be predicted and the text descriptions corresponding to the relations in the triplet to be predicted; the encoder output is the last hidden state of the first BERT model after L2 regularization and average pooling, which is used as the output to obtain the head entity relation-aware embedding;

[0062] The second BERT model computes the L2-regularized embeddings of the coarsely filtered candidate tail entity set. The input to the second BERT model can include the text descriptions of the coarsely filtered candidate tail entities; the encoder output is the L2-regularized last hidden state of the second BERT model, which is the candidate tail entity embedding.

[0063] In some embodiments, the triplet prediction module can use the dot product to calculate the similarity between the head entity relation-aware embedding and the entity embedding as a score for the coarsely filtered candidate entities.

[0064] Step 260: Based on the scores of the candidate tail entities, determine the tail entities included in the triples to be predicted. In some embodiments, step 260 may be performed by the triple prediction module.

[0065] In some embodiments, the triplet prediction module may use the candidate entity with the highest score as the final prediction result.

[0066] It should be noted that the above description of the knowledge graph completion method based on multi-view link prediction is for illustrative purposes only and does not limit the scope of this specification. Those skilled in the art can make various modifications and changes to the knowledge graph completion method based on multi-view link prediction under the guidance of this specification. However, these modifications and changes are still within the scope of this specification.

[0067] In some embodiments, the knowledge graph completion method and system based on multi-view link prediction have at least the following beneficial effects:

[0068] 1. In order to ensure that the constructed negative samples are high-quality negative samples, the constructed negative samples must satisfy the existing knowledge concepts of the corresponding entities.

[0069] 2. To improve the uncertainty of link prediction, we consider establishing a multi-view link prediction model by introducing existing knowledge concepts. First, we use existing knowledge concepts to perform a coarse-grained filtering of candidate entities, and then we use similarity to further refine the discrimination of candidate entities that meet the existing knowledge concepts, thereby improving the accuracy of knowledge graph completion.

[0070] The basic concepts have been described above. Obviously, for those skilled in the art, the detailed disclosure above is merely illustrative and does not constitute a limitation of this specification. Although not explicitly stated herein, those skilled in the art may make various modifications, improvements, and corrections to this specification. Such modifications, improvements, and corrections are suggested in this specification and therefore remain within the spirit and scope of the exemplary embodiments described herein.

[0071] Furthermore, this specification uses specific terms to describe embodiments thereof. For example, "an embodiment," "one embodiment," and / or "some embodiments" refer to a particular feature, structure, or characteristic associated with at least one embodiment of this specification. Therefore, it should be emphasized and noted that references to "an embodiment," "one embodiment," or "an alternative embodiment" in different locations throughout this specification do not necessarily refer to the same embodiment. Moreover, certain features, structures, or characteristics in one or more embodiments of this specification can be appropriately combined.

[0072] Furthermore, unless expressly stated in the claims, the order of processing elements and sequences, the use of numbers and letters, or other names described in this specification are not intended to limit the order of the processes and methods described herein. Although various examples have been discussed in the foregoing disclosure of some embodiments of the invention that are currently considered useful, it should be understood that such details are for illustrative purposes only, and the appended claims are not limited to the disclosed embodiments; rather, the claims are intended to cover all modifications and equivalent combinations that conform to the spirit and scope of the embodiments described herein. For example, while the system components described above can be implemented using hardware devices, they can also be implemented solely using software solutions, such as installing the described system on existing servers or mobile devices.

[0073] Similarly, it should be noted that, in order to simplify the description disclosed herein and thus aid in the understanding of one or more embodiments of the invention, the foregoing description of embodiments in this specification may sometimes combine multiple features into a single embodiment, drawing, or description thereof. However, this method of disclosure does not imply that the subject matter of this specification requires more features than those mentioned in the claims. In fact, the embodiments contain fewer features than all the features of a single embodiment disclosed above.

[0074] Finally, it should be understood that the embodiments described in this specification are merely illustrative of the principles of the embodiments described herein. Other variations may also fall within the scope of this specification. Therefore, alternative configurations of the embodiments described herein are intended to be illustrative rather than limiting, and should be considered consistent with the teachings of this specification. Accordingly, the embodiments described herein are not limited to those explicitly introduced and described herein.

Claims

1. A knowledge graph completion method based on multi-view link prediction, characterized in that, include: Establish an automatic converter for existing knowledge concepts; Build and train a dual encoder; Obtain the triplet to be predicted; The existing knowledge concept automatic converter determines the candidate tail entity and candidate tail entity text description corresponding to the triple to be predicted based on the head entity and relations included in the triple to be predicted. The dual encoder determines the score of the candidate tail entity based on the candidate tail entity and the candidate tail entity text description; Based on the scores of the candidate tail entities, the tail entities included in the triplet to be predicted are determined; The establishment and training of the dual encoder includes: An initial dual encoder is established, wherein the initial dual encoder includes a first BERT model and a second BERT model; The initial dual encoder is trained using positive and negative samples to generate the dual encoder. The first BERT model determines the head entity relationship-aware embedding based on the head entity text description corresponding to the head entity in the triplet to be predicted and the relationship text description corresponding to the relationship. The second BERT model determines the candidate tail entity embedding based on the candidate tail entity text description; The score of the candidate tail entity is determined based on the similarity between the head entity relation-aware embedding and the candidate tail entity embedding.

2. The knowledge graph completion method based on multi-view link prediction according to claim 1, characterized in that, The establishment of an automatic converter for existing knowledge concepts includes: Establish a dictionary mapping entities to existing knowledge concepts and a dictionary mapping entities to textual descriptions. By using the entity-to-existing knowledge concept mapping dictionary and the entity-to-text description mapping dictionary, the triples in the triple dataset are transformed into existing knowledge concept triples, and an existing knowledge concept triple dataset corresponding to the triple dataset is established.

3. The knowledge graph completion method based on multi-view link prediction according to claim 1, characterized in that, Obtaining the negative sample includes: Generate hard negative samples and soft negative samples; Based on existing knowledge, the soft negative samples are screened.

4. The knowledge graph completion method based on multi-view link prediction according to claim 1, characterized in that, The existing knowledge concept automatic converter determines the candidate tail entities and candidate tail entity text descriptions corresponding to the triples to be predicted based on the head entities and relations included in the triples to be predicted, including: Based on the set of existing knowledge concepts corresponding to the head entities included in the triplet to be predicted and the relationships included in the triplet to be predicted, the set of candidate existing knowledge concepts corresponding to the tail entities included in the triplet to be predicted is determined. The entity-to-entity-text-description mapping dictionary determines the candidate tail entity and candidate tail entity text description corresponding to the triplet to be predicted based on the set of candidate existing knowledge concepts corresponding to the tail entity.

5. A knowledge graph completion system based on multi-view link prediction, used to execute the knowledge graph completion method based on multi-view link prediction as described in claim 1, characterized in that, include: The converter creation module is used to create automatic converters for existing knowledge concepts; The encoder creation module is used to create and train dual encoders; The triplet acquisition module is used to acquire triplets to be predicted. The triplet prediction module is used to determine the candidate tail entities and candidate tail entity text descriptions corresponding to the triplet to be predicted based on the head entities and relationships included in the triplet to be predicted using the existing knowledge concept automatic converter. The triplet prediction module is also used to determine the score of the candidate tail entity based on the candidate tail entity and the candidate tail entity text description using the dual encoder; The triplet prediction module is also used to determine the tail entities included in the triplet to be predicted based on the scores of the candidate tail entities.

6. A knowledge graph completion system based on multi-view link prediction according to claim 5, characterized in that, The converter establishment module is also used for: Establish a dictionary mapping entities to existing knowledge concepts and a dictionary mapping entities to textual descriptions. By using the entity-to-existing knowledge concept mapping dictionary and the entity-to-text description mapping dictionary, the triples in the triple dataset are transformed into existing knowledge concept triples, and an existing knowledge concept triple dataset corresponding to the triple dataset is established.

Images