A knowledge graph embedding method based on hyperbolic space-based join entity description
By constructing a knowledge graph embedding model in hyperbolic space and combining it with entity description information, the problem of existing methods failing to effectively utilize entity descriptions is solved, achieving better hierarchical structure representation and semantic relationship capture, thus improving the embedding effect of knowledge graphs.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- INNER MONGOLIA UNIVERSITY
- Filing Date
- 2023-08-11
- Publication Date
- 2026-05-19
AI Technical Summary
Existing hyperbolic space knowledge graph embedding methods fail to effectively utilize entity description information, thus affecting the embedding results.
A knowledge graph embedding model is constructed in hyperbolic space. Combining word embedding and graph embedding models, entity text embedding vectors are processed by Einstein midpoint pooling, and the Poincaré sphere model is optimized using scoring and loss functions to achieve entity and relation embedding.
It improves the hierarchical structure representation capability of knowledge graphs, captures hierarchical and semantic relationships between entities, enriches entity representation capabilities, alleviates data sparsity problems, and supports more complex reasoning capabilities.
Smart Images
Figure CN116992049B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of knowledge graph representation technology, specifically relating to a knowledge graph embedding method based on hyperbolic space and incorporating entity descriptions. Background Technology
[0002] Knowledge graphs (KGs), as a structured knowledge representation method containing a large number of fact triples (head entity, relation, tail entity), have been widely applied in many fields, such as information retrieval, recommender systems, service computing, and question answering systems, and have shown great potential. However, real-world knowledge graphs, such as FreeBase, WordNet, Yago, and DBPedia, are often highly incomplete, and manual completion would be extremely costly. Therefore, knowledge graph completion based on graph embedding representation methods has attracted considerable attention due to its low cost and high efficiency.
[0003] Hyperbolic space can be considered a continuous simulation of trees, exhibiting minimal distortion and being particularly well-suited for hierarchical structures. This is because hyperbolic space closely matches the growth rate of tree-like data, unlike Euclidean space. Furthermore, hyperbolic space provides high-quality representations even at low embedding dimensions, making it highly suitable for modeling hierarchical data. In other words, hyperbolic geometry offers embedding models greater possibilities and a wider operational space to fully explore and reproduce the underlying structure and fundamental features of knowledge graphs. Currently, some hyperbolic embedding methods have achieved excellent results on knowledge graph link prediction tasks, even surpassing Euclidean embedding methods.
[0004] Knowledge graph embedding methods based on Euclidean space have a fundamental limitation: their ability to model complex patterns is inherently limited by the dimension of the embedding space. In Euclidean space, the volume of a sphere grows only polynomially with respect to its radius, which leads to highly distorted embeddings.
[0005] Current knowledge graph embedding methods in hyperbolic spaces only consider the graph structure and fail to utilize additional information containing rich features. However, besides graph structure information, knowledge graphs, as a special type of heterogeneous graph, allow each entity and relation to be associated with specific textual concepts or category information. In Euclidean space, graph embedding methods that introduce additional information supplement the lack of structural information through the textual knowledge and hierarchical information inherent in the knowledge graph itself, thereby learning more effective graph embeddings that better model the knowledge graph. However, currently, there are very few knowledge graph embedding methods in the hyperbolic space domain that incorporate additional information such as entity descriptions. Summary of the Invention
[0006] To address the aforementioned shortcomings in existing technologies, the knowledge graph embedding method based on hyperbolic space and incorporating entity descriptions provided by this invention solves the problem that hyperbolic space-based knowledge graph embedding methods do not consider entity descriptions, thus affecting the knowledge graph embedding effect.
[0007] To achieve the aforementioned objectives, the present invention employs the following technical solution: a knowledge graph embedding method based on hyperbolic space with added entity descriptions, comprising the following steps:
[0008] S1. Construct a knowledge graph;
[0009] S2. Construct a hyperbolic embedding model corresponding to the knowledge graph, incorporating entity descriptions;
[0010] The hyperbolic embedding model includes a word embedding model and a graph embedding model;
[0011] S3. In hyperbolic space, obtain entity text embedding vectors using a word embedding model;
[0012] S4. Perform Einstein midpoint pooling on the entity text embedding vector.
[0013] S5. Use the pooled entity text embedding vector as the initial embedding vector of the entity in the graph embedding model, and randomly initialize the embedding vector of the relationship between the entities to obtain the initial vectors of the entities and relationships in the Poincaré sphere model.
[0014] S6. Based on the initial vectors of entities and relations, the graph structure of the knowledge graph is embedded into the Poincaré sphere model through a graph embedding model. The Poincaré sphere model is then trained and optimized using a scoring function and a loss function to achieve knowledge graph embedding.
[0015] Furthermore, in step S2, the word embedding model is a Glove word embedding model trained in the Poincaré sphere model;
[0016] The specific method for obtaining entity description text vectors using the GloVe word embedding model is as follows:
[0017] S31. Train the Glove model in the Poincaré sphere model to obtain word embedding vectors in the Poincaré sphere model;
[0018] S32. In hyperbolic space, based on entity descriptions in the knowledge graph and combined with word embedding vectors, determine the entity text embedding vector of each entity in the Poincaré sphere model.
[0019] Further, in step S31, the loss function for training the GloVe model is:
[0020]
[0021] In the formula, This is the loss value. X is the size of the vocabulary. ij Let f(·) be the number of times word i and its context word j appear in the same window context in the corpus, and let f(·) be the weight function. For the distance formula in the Poincaré sphere model, w i Let i be the embedding vector of word i. Let b be the embedding vector of the context word j. i For the bias term of the word 'i', This is the bias term for the context word j.
[0022] Furthermore, in step S32, for a given entity e in the knowledge graph... i Description text of ∈ε make This represents the word "w" in the entity description text. i The embedding vectors are used to obtain the entity text embedding vectors for each entity in the Poincaré sphere model. for;
[0023]
[0024] In the formula, To represent the Poincaré sphere model, i is the word ordinal number.
[0025] Further, step S4 specifically includes:
[0026] S41. Project the entity text embedding vectors in the Poincaré sphere model onto the Klein model and calculate the Einstein midpoint;
[0027] S42. Based on the conversion formula between the Poincaré sphere model and the Klein model, project the Einstein midpoint back to the Poincaré sphere model;
[0028] S43. Based on the Einstein midpoint in the Poincaré sphere model, calculate the weighted average of the entity text embedding vectors to obtain the pooled entity text embedding vectors.
[0029] Furthermore, in step S42, the conversion formula between the Poincaré sphere model and the Klein model is:
[0030]
[0031]
[0032] In the formula, and These are the embedding vectors in the Poincaré sphere model and the Klein model, respectively. and , respectively, are the Einstein midpoint pooling vectors in the Poincaré sphere model and the Klein model, where c is the curvature.
[0033] Furthermore, in step S43, the formula for calculating the entity text embedding vector after pooling is:
[0034]
[0035] In the formula, This is the Einstein midpoint pooling vector of the entity text in the Klein model, i.e., the entity text embedding vector after pooling. x is the Klein factor. i Embed vectors for entity text. This is the Klein model.
[0036] Furthermore, in step S6, the scoring function is:
[0037]
[0038] In the formula, e i For the head entity, R j For the relationship, E k For tail entities, This is the distance function in the Poincaré sphere model. This is the embedding vector of the head entity after transformation through a specific relation. The tail entity is the embedding vector after transformation through a specific relation, R is the diagonal relation matrix, and h is the embedding vector. i For hyperbolic embedding of the head entity, h k Hyperbolic embedding for tail entity, For Möbius addition, r h For relation r j The hyperbolic translation vector, b i For the bias term of the head entity, b k For the bias term of the tail entity, where h i , b i , It is a Poincaré sphere model.
[0039] Furthermore, in step S6, the loss function for training the Poincaré sphere model is:
[0040]
[0041] In the formula, It is the predicted probability value. It is the binary label of a positive or negative sample. For the i-th probability prediction value, The binary label is the i-th positive or negative sample.
[0042] Further, in step S6, the Poincaré sphere model is optimized using Riemann stochastic gradient descent, where the update formula for the Riemann gradient is:
[0043]
[0044] In the formula, η is the learning rate. It is the Euclidean gradient multiplied by the inverse of the Poincaré metric tensor.
[0045] The beneficial effects of this invention are as follows:
[0046] (1) The method of the present invention has better hierarchical structure representation ability and can capture the hierarchical relationship and semantic hierarchical relationship between entities, which is very beneficial for reasoning and inference tasks of knowledge graph.
[0047] (2) By adding entity descriptions, the method of the present invention can enrich the representation of entities, provide more comprehensive semantic information, help alleviate the problem of data sparsity, enable the embedding model to better handle missing information, and help the embedding model to better understand the semantic relationship between entities, thereby supporting more complex reasoning capabilities.
[0048] (3) The content of entity descriptions can be designed and expanded according to specific application scenarios, thereby improving the adaptability of knowledge graph embedding methods. Attached Figure Description
[0049] Figure 1 The flowchart of the knowledge graph embedding method based on hyperbolic space and incorporating entity descriptions provided by the present invention is shown.
[0050] Figure 2 This is a schematic diagram of the hyperbolic embedding model provided by the present invention. Detailed Implementation
[0051] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.
[0052] This invention provides a knowledge graph embedding method based on hyperbolic space and incorporating entity descriptions, such as... Figure 1 As shown, it includes the following steps:
[0053] S1. Construct a knowledge graph;
[0054] S2. Construct a hyperbolic embedding model corresponding to the knowledge graph, incorporating entity descriptions;
[0055] The hyperbolic embedding model includes a word embedding model and a graph embedding model;
[0056] S3. In hyperbolic space, obtain entity text embedding vectors using a word embedding model;
[0057] S4. Perform Einstein midpoint pooling on the entity text embedding vector.
[0058] S5. Use the pooled entity text embedding vector as the initial embedding vector of the entity in the graph embedding model, and randomly initialize the embedding vector of the relationship between the entities to obtain the initial vectors of the entities and relationships in the Poincaré sphere model.
[0059] S6. Based on the initial vectors of entities and relations, the graph structure of the knowledge graph is embedded into the Poincaré sphere model through a graph embedding model. The Poincaré sphere model is then trained and optimized using a scoring function and a loss function to achieve knowledge graph embedding.
[0060] In step S1 of this embodiment of the invention, the constructed knowledge graph is represented as follows: Where ε is the set of entities, For a set of relations, For a set of triples, A collection of text describing entities; Each triple in the array is specifically represented as (e i ,r j ,e k ), where e i ∈ε is the head entity of the triple, e k ∈ε is the tail entity. It is a relation type. For each entity e i Each ∈ε has a descriptive text. Where w i It is a vocabulary list The words in the text.
[0061] Knowledge graphs are generally incomplete, so a link prediction task is used to optimize the embedding of entities and relationships in the graph and to infer and complete the missing parts. In this embodiment, the following methods are used respectively: Represents entity e i ∈ε and relation The embedding vectors in the hyperbolic space Poincaré ball model are given by d, where d is the dimension of the embedding space and c is the curvature. Typically, for each triplet, an embedding vector is obtained through a scoring function. Learning a fraction s = φ(e i ,r j ,e k ), representing the predictive strength of the current triplet's true reality. Through continuous learning, the positive sample triplet is improved. The scores are getting higher and higher, negative samples The scores kept decreasing. Then, the scores were transformed into predicted probabilities of the triplet being true using a non-linear function.
[0062] In step S2 of this embodiment of the invention, the framework of the constructed hyperbolic embedding model HyEED is as follows: Figure 2 As shown, from an overall structural perspective, the model can be divided into two parts: a word embedding model and a graph embedding model. The word embedding model aims to obtain text vectors describing entities within a Poincaré sphere in hyperbolic space, while the graph embedding model aims to embed the graph structure of the knowledge graph into the Poincaré sphere for better modeling. Figure 2 In this context, Einstein midpoint Pooling; Random Embedding; Score; Mission: Impossible III is a 2006 American action spy film… and J.A. Brams is an American film and television producer… represent entity description text; PoincaréGloVe is a word embedding model.
[0063] In step S2 of this embodiment, since the ultimate goal is to embed the knowledge graph into the Poincaré sphere in hyperbolic space, the word embedding model in this embodiment is the Glove word embedding model (i.e., the PoincaréGlove model) trained in the Poincaré sphere model. Glove is a word representation tool based on global word frequency statistics. It learns word representations in Euclidean space by statistically analyzing the occurrence of words in a text corpus, with the aim of capturing the meaning and relationships of words from a geometric perspective.
[0064] Based on this, in step S3 of this embodiment, the method for obtaining entity description text vectors using the GloVe word embedding model is specifically as follows:
[0065] S31. Train the Glove model in the Poincaré sphere model to obtain word embedding vectors in the Poincaré sphere model;
[0066] S32. In hyperbolic space, based on entity descriptions in the knowledge graph and combined with word embedding vectors, determine the entity text embedding vector of each entity in the Poincaré sphere model.
[0067] Specifically, in step S31, in order to obtain word embedding vectors in hyperbolic space, a GloVe model is trained in the Poincaré sphere using a Wikipedia corpus containing 1.4 billion English sentences. However, training the PoincaréGlove model faces a challenge: there is no explicit function definition in hyperbolic space like the Euclidean inner product, so the loss function of the GloVe model shown below cannot be used directly.
[0068]
[0069] Replacing the inner product in the original GloVe loss function above with the distance within the Poincaré sphere, we obtain the following loss function for training the GloVe model:
[0070]
[0071] In the formula, This is the loss value. X is the size of the vocabulary. ij Let f(·) be the number of times word i and its context word j appear in the same window context in the corpus, and let f(·) be the weight function. For the distance formula in the Poincaré sphere model, w i Let i be the embedding vector of word i. Let b be the embedding vector of the context word j. i For the bias term of the word 'i', This is the bias term for the context word j.
[0072] In step S32, after training the GloVe model, word embedding vectors in the Poincaré sphere can be obtained for a given entity e in the knowledge graph. i Description text of ∈ε make This represents the word "w" in the entity description text. i The embedding vectors are used to obtain the entity text embedding vectors for each entity in the Poincaré sphere model. for;
[0073]
[0074] In the formula, To represent the Poincaré sphere model, i is the word ordinal number.
[0075] Based on obtaining the entity text embedding vectors, the relationships between entities are... Embedded vector Random initialization is used to obtain the text vectors of entities and the vectors of relationships in the Poincaré sphere in hyperbolic space. However, the text vectors of entities cannot be directly used as the initial vectors of entities in the graph embedding model. They need to be further processed. In this embodiment, pooling technology is used to reduce the dimensionality of the output of the word embedding model and convert it into an input form suitable for the graph embedding model.
[0076] Based on this, step S4 of this embodiment of the invention is specifically as follows:
[0077] S41. Project the entity text embedding vectors in the Poincaré sphere model onto the Klein model and calculate the Einstein midpoint;
[0078] S42. Based on the conversion formula between the Poincaré sphere model and the Klein model, project the Einstein midpoint back to the Poincaré sphere model;
[0079] S43. Based on the Einstein midpoint in the Poincaré sphere model, calculate the weighted average of the entity text embedding vectors to obtain the pooled entity text embedding vectors.
[0080] In step S42 above, the conversion formula between the Poincaré sphere model and the Klein model is:
[0081]
[0082]
[0083] In the formula, and These are the embedding vectors in the Poincaré sphere model and the Klein model, respectively. and , respectively, are the Einstein midpoint pooling vectors in the Poincaré sphere model and the Klein model, where c is the curvature.
[0084] In step S43 above, the formula for calculating the entity text embedding vector after pooling is:
[0085]
[0086] In the formula, This is the Einstein midpoint pooling vector of the entity text in the Klein model, i.e., the entity text embedding vector after pooling. x is the Klein factor. i Embed vectors for entity text. This is the Klein model.
[0087] Specifically, in Natural Language Processing (NLP), average pooling can downsample the input data, thereby reducing the number of model parameters and computational cost; in Euclidean space, average pooling is as follows:
[0088]
[0089] Since the embedding is performed in hyperbolic space in this embodiment, a weighted midpoint method called Einstein midpoint is used to extend average pooling to hyperbolic space. This is relevant to the d-dimensional Klein model. In this embodiment, the weighted average of the embedding vectors obtained through the Einstein midpoint is calculated. However, this embodiment uses the Poincaré sphere model in the embedding space instead of the Klein model. Therefore, the above formula cannot be used directly to calculate the Einstein midpoint. Since they are isomorphic and equivalent, the entity description text vector can be projected onto the Klein model first to calculate the Einstein midpoint, and then the result can be projected back to the Poincaré model according to the conversion formula between the two models.
[0090] The entity text embedding vector processed by Einstein midpoint pooling can be used as the initial vector of the entity in the graph embedding model. However, in many other knowledge graph embedding models, the initial vector of the entity is generally randomly initialized. Einstein midpoint pooling technology is the key point in this embodiment to solve how to link and combine entity description text information and graph embedding information.
[0091] In step S6 of this embodiment of the invention, the scoring function is:
[0092]
[0093] In the formula, e i For the head entity, R j For the relationship, E k For tail entities, This is the distance function in the Poincaré sphere model. This is the embedding vector of the head entity after transformation through a specific relation. The embedding vector is the tail entity transformed through a specific relation, r is the diagonal relation matrix, and h is the embedding vector. i For hyperbolic embedding of the head entity, h k Hyperbolic embedding for tail entity, For Möbius addition, r H For relation r j The hyperbolic translation vector, b i For the bias term of the head entity, b k For the bias term of the tail entity, where h i , It is a Poincaré sphere model.
[0094] Specifically, the process of calculating the probability of each triplet being true based on the above scoring function is as follows:
[0095] First of all pass Project onto the tangent space of the Poincaré sphere at point 0, then multiply by the diagonal relation matrix. Finally passed Project back into the Poincaré sphere. Depend on and The scores are calculated using the Möbius method. After obtaining the scores, the probability of each triplet being true is calculated using the logistic sigmoid function.
[0096] In step S6 of this invention, during the training of the Poincaré sphere model, data augmentation techniques are used for each triple in terms of the dataset. Generate an inverse triplet This doubles the training data. When training the model, for each positive triplet (E... i ,r jE k Generate k negative triples (negative samples) or é i and é k It is randomly selected from the entity set ε, where the loss function for training the Poincaré sphere model is:
[0097]
[0098] In the formula, It is the predicted probability value. It is the binary label of a positive or negative sample. For the i-th probability prediction value, The binary label is the i-th positive or negative sample.
[0099] In step S6 of this embodiment of the invention, during the optimization of the Poincaré sphere model, Riemann stochastic gradient descent (RSGD) is used for optimization. To calculate the Riemann gradient... The Euclidean gradient needs to be calculated. Multiply by the inverse of the Poincaré metric tensor, i.e.
[0100] When updating the gradient, since it is embedded in hyperbolic space, the Euclidean gradient update step is not used directly. Instead, it uses a first-order approximation updated by the true Riemannian. The specific steps are as follows: Exponential operations will use the Riemann gradient Projecting the geodesic onto the Poincaré sphere, where the geodesic represents the shortest path between two points in hyperbolic space, and then updating the Riemann gradient according to the following formula:
[0101]
[0102] In the formula, η is the learning rate.
[0103] The knowledge graph embedding method proposed in this invention incorporates entity description text. In hyperbolic space, it uses the PoincaréGlove model to extract rich semantic information from the entity description text that supplements the graph structure, and then uses the Einstein midpoint pooling method to fuse textual and graph structure information. Compared to current baseline models, this invention's model achieves better performance on knowledge graph link prediction tasks. Furthermore, adding entity description information in hyperbolic space does indeed improve the performance of graph embedding methods. For knowledge graph data with unclear hierarchical structures, adding entity description data can still compensate for the shortcomings of embedding non-hierarchical data in hyperbolic space and reduce its negative impact.
[0104] Specific embodiments have been used to illustrate the principles and implementation methods of this invention. The descriptions of the embodiments above are only for the purpose of helping to understand the method and core ideas of this invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.
[0105] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the spirit of the invention, and these modifications and combinations are still within the scope of protection of this invention.
Claims
1. A knowledge graph embedding method based on hyperbolic space and incorporating entity descriptions, executed by a computer device, wherein the method is stored in a computer-readable storage medium and executed by a processor, and is used to perform link prediction and automatic completion processing on knowledge graph data stored in a computer system, characterized in that, Includes the following steps: S1. Obtain knowledge graph data containing entity sets, relation sets, triple sets, and entity description text sets; S2. Construct a hyperbolic embedding model corresponding to the knowledge graph, incorporating entity descriptions; The hyperbolic embedding model includes a word embedding model and a graph embedding model; In step S2, the word embedding model is the Glove word embedding model trained in the Poincaré sphere model; S3. In hyperbolic space, obtain entity text embedding vectors using a word embedding model; The specific method for obtaining entity description text vectors using the GloVe word embedding model is as follows: S31. Train the GloVe model in the Poincaré sphere model to obtain the word embedding vectors in the Poincaré sphere model; S32. In hyperbolic space, based on entity descriptions in the knowledge graph and combined with word embedding vectors, determine the entity text embedding vector of each entity in the Poincaré sphere model; S4. Perform Einstein midpoint pooling on the entity text embedding vector, specifically: S41. Project the entity text embedding vectors in the Poincaré sphere model onto the Klein model and calculate the Einstein midpoint; S42. Based on the conversion formula between the Poincaré sphere model and the Klein model, project the Einstein midpoint back to the Poincaré sphere model; S43. Based on the Einstein midpoint in the Poincaré sphere model, calculate the weighted average of the entity text embedding vectors to obtain the pooled entity text embedding vectors. S5. Use the pooled entity text embedding vector as the initial embedding vector of the entity in the graph embedding model, and randomly initialize the embedding vector of the relationship between the entities to obtain the initial vectors of the entities and relationships in the Poincaré sphere model. S6. Based on the initial vectors of entities and relations, the graph structure of the knowledge graph is embedded into the Poincaré sphere model through the graph embedding model, and the Poincaré sphere model is trained and optimized using the scoring function and loss function. S7. The optimized Poincaré sphere model, which includes optimized entity and relation embedding vectors, is used for the knowledge graph link prediction task to predict missing triples and complete the knowledge graph embedding.
2. The knowledge graph embedding method based on hyperbolic space and incorporating entity descriptions according to claim 1, characterized in that, In step S31, the loss function for training the GloVe model is: In the formula, The loss value. For vocabulary size, For words in the corpus i and its context words j The number of times it appears in the same window context. For the weight function, This is the distance formula in the Poincaré sphere model. For words i Embedded vector, for context words Embedded vector, For words The bias term, for context words j The bias term.
3. The knowledge graph embedding method based on hyperbolic space and incorporating entity descriptions according to claim 2, characterized in that, In step S32, for a given entity in the knowledge graph Description text ,make Represents words in entity description text The embedding vectors are used to obtain the entity text embedding vectors for each entity in the Poincaré sphere model. for; In the formula, To represent the Poincaré sphere model. i For word ordinal numbers.
4. The knowledge graph embedding method based on hyperbolic space and incorporating entity descriptions according to claim 1, characterized in that, In step S42, the conversion formula between the Poincaré sphere model and the Klein model is as follows: In the formula, and These are the embedding vectors in the Poincaré sphere model and the Klein model, respectively. and These are the Einstein midpoint pooling vectors in the Poincaré sphere model and the Klein model, respectively. For curvature.
5. The knowledge graph embedding method based on hyperbolic space and incorporating entity descriptions according to claim 4, characterized in that, In step S43, the formula for calculating the entity text embedding vector after pooling is: In the formula, This is the Einstein midpoint pooling vector of the entity text in the Klein model, i.e., the entity text embedding vector after pooling. Klein factor, Embed vectors for entity text. This is the Klein model.
6. The knowledge graph embedding method based on hyperbolic space and incorporating entity descriptions according to claim 4, characterized in that, In step S6, the scoring function is: In the formula, For the head entity, For the relationship, For tail entities, This is the distance function in the Poincaré sphere model. This is the embedding vector of the head entity after transformation through a specific relation. The embedding vector of the tail entity after transformation through a specific relation. It is a diagonal relation matrix. Hyperbolic embedding for head entity, Hyperbolic embedding for tail entity, For Möbius addition, For relationship hyperbolic translation vector, For the bias term of the head entity, For the bias term of the tail entity, where, , , , , , , , It is a Poincaré sphere model.
7. The knowledge graph embedding method based on hyperbolic space and incorporating entity descriptions according to claim 4, characterized in that, In step S6, the loss function for training the Poincaré sphere model is: In the formula, It is the predicted probability value. It is the binary label of a positive or negative sample. For the first i One probability prediction value, For the first i A binary label for a positive or negative sample.
8. The knowledge graph embedding method based on hyperbolic space and incorporating entity descriptions according to claim 4, characterized in that, In step S6, the Poincaré sphere model is optimized using Riemann stochastic gradient descent, where the update formula for the Riemann gradient is: In the formula, For learning rate, It is the Euclidean gradient multiplied by the inverse of the Poincaré metric tensor.