An entity alignment method based on noise optimization and attribute fusion
Through the method of noise optimization and attribute fusion, the noise perception generator and discriminator iteratively obtain pure seed entity pairs, and embed them in combination with GCN and pretrained word vectors, solving the error accumulation problem caused by the breadth of information sources and labeling differences in entity alignment, and improving the accuracy and effect of entity alignment.
Patent Information
- Application Number
- CN202310720797.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-06-19
- Publication Date
- 2025-08-26
- Estimated Expiration
- 2043-06-19
AI Technical Summary
In the process of entity alignment, the prior art has accumulated errors due to the widespread information source and the difference in manual labeling, and ignores the contribution of attribute information in the knowledge graph, resulting in poor entity alignment.
Using a method based on noise optimization and attribute fusion, pure seed entity pairs are iteratively obtained through noise perception optimization generator and discriminator, and structure and attribute embedding are used using GCN and pretrained word vectors, denoising is performed in combination with high-speed gating mechanism, and finally the structure alignment model is obtained.
It improves the accuracy and effectiveness of entity alignment, reduces complexity and experimental overhead, and enhances the driving role of entity alignment tasks.
Smart Images
Figure CN116737957B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of knowledge graph technology, and in particular relates to an entity alignment method based on noise optimization and attribute fusion. Background Art
[0002] With the advancement of science and technology, the amount of information coming from all walks of life is becoming increasingly vast. How to store this information and capture the complex relationships between it has become crucial. The rate at which information is growing far exceeds the speed of human comprehension. It can be said that human life is surrounded by an overwhelming amount of information. As a powerful and comprehensive information visualization platform encompassing knowledge from diverse fields, knowledge graphs play a key role in the integration of various disciplines. Knowledge fusion aims to align and merge heterogeneous and redundant information within knowledge graphs to form globally unified knowledge identifiers and associations. To further promote knowledge fusion, researchers have focused on knowledge graph-based entity alignment, which is one of the most fundamental and critical tools in the knowledge fusion process. Therefore, we need to develop an entity alignment model that meets societal needs.
[0003] The birth of the semantic web is based on the idea of data linking and the refinement of relevant technical standards, which subsequently led to the emergence of knowledge graphs. In November 2012, Google first proposed the concept of knowledge graphs, stating that adding knowledge graph functionality would enhance search. Information is relevant, shareable, and transferable, and the desired knowledge graph needs to span various fields. To facilitate search and query, knowledge aggregation is essential, enabling efficient coordination between large-scale knowledge graphs. However, the sheer volume of information makes it impossible for a single entity to construct a relatively complete and accurate knowledge graph. In other words, different institutions and organizations are working on this information aggregation and using different construction methods or logical design approaches to build unique knowledge graphs. This ultimately leads to heterogeneity and redundancy among different knowledge graphs, and manually annotated information also contains certain errors.
[0004] There are usually two forms of entity alignment: one is the relationship triple represented by "entity-relationship-entity", and the other is the attribute triple represented by "entity-attribute-entity". Among them, attribute information is of great benefit to the effect of entity alignment. Existing research on entity alignment of knowledge graphs can be divided into two categories: (1) traditional entity alignment methods, such as entity alignment methods based on similarity calculation and entity alignment methods based on relational reasoning; (2) entity alignment methods based on representation learning, such as entity alignment methods based on translation models, entity alignment methods based on graph neural network models, entity alignment research based on multi-angle fusion, and entity alignment methods based on deep models.
[0005] The development of the data age has prompted people to organize and study massive amounts of information. In the field of natural language processing, entity alignment based on knowledge graphs has become a growing research hotspot among experts and scholars due to its outstanding contributions. Because various institutions have their own standards and specifications when constructing knowledge graphs, the resulting knowledge graphs are heterogeneous. While most studies assume that labeled and aligned seed entity pairs are completely correct and can be used directly for data training, this often leads to the accumulation of errors in subsequent entity alignment tasks, reducing the effectiveness of entity alignment.
[0006] Currently, a large number of studies employ supervised or semi-supervised approaches to entity alignment. For example, after aligning entities using a manually annotated seed set, new entity matching pairs are then added to the original seed set for training, forming a semi-supervised approach to expand the seed set. On the one hand, the wide range of information sources and the variability of manually annotated entities can lead to the accumulation of errors in the entity alignment task, which to some extent affects the effectiveness of entity alignment.
[0007] On the other hand, the entity alignment research method considers a single structural embedding, ignoring the favorable external condition that the knowledge graph contains attribute information. Summary of the Invention
[0008] The purpose of the embodiments of the present invention is to provide an entity alignment method based on noise optimization and attribute fusion, aiming to solve the problems raised in the above background technology.
[0009] The embodiment of the present invention is implemented as follows: an entity alignment method based on noise optimization and attribute fusion includes the following steps:
[0010] Step 1: Initialize the entity seed set with noise to obtain accurate noise entity pairs;
[0011] Step 2: Inject the noise-aware optimization generator to obtain more noise entity pairs around the relevant distribution of the true seed entity pairs;
[0012] Step 3: After the noise entity pairs are generated, they are fed into the noise-aware optimization discriminator to identify the true seed entity pairs and the noise entity pairs;
[0013] Step 4: The noise-aware optimized generator and the noise-aware optimized discriminator iterate each other to obtain a trust score and compare it with the threshold θ to obtain clean labeled data;
[0014] Step 5: Use GCN to embed the obtained real seed entity pairs into structures and finally obtain the structural alignment model.
[0015] A further technical solution is that in step 1, in the initial state, in order to enable the model to accurately determine the noise entity and provide initialization data for the identification model, thereby improving the probability of noise extraction, it is necessary to artificially generate accurate noise entity pairs in advance; extract a certain proportion of seed entity pairs from the seed set containing the label and randomly replace them, and after the replacement is completed, merge them with the remaining entity pairs to obtain accurate noise entity pairs.
[0016] A further technical solution is that in step 2, after initialization, the noise-aware optimization generator acquires noise entities around the distribution of true seed entity pairs to obtain more noise entity pairs. This acquisition process ensures that the noise entities are as close as possible to the true seed entity pairs to deceive the noise-optimized discriminator model. The probabilities of the entity pairs generated by the noise-aware optimization generator are based on negative samples generated using Bernoulli negative sampling.
[0017] A further technical solution is that in step 4, during the model training phase, the correct seed entity pair is determined by the threshold θ. When the classification function is greater than or equal to the threshold, the score function value is recorded as 1, that is, the classification function sigmoid (f τ (e i , e j ))≥θ, the score function Score(e i , e j ) is recorded as 1; that is, the current identification entity is considered to be the correct entity. On the contrary, when the classification function is less than the threshold, the record score function value is 0, that is, the classification function is recorded as sigmoid (f τ (e i , e j ))<θ, score function Score(e i , e j ) is recorded as 0; the number of correct entity seeds is continuously expanded, the entity pairs generated by the noise-aware optimization generator are incorrect entity pairs, and the noise-aware optimization discriminator evolves into a two-classification problem to distinguish positive and negative samples.
[0018] A further technical solution is to use pre-trained word vectors as the input for structural embedding in step 5. Compared with traditional structural information conversion and encoding, pre-trained word vectors contain rich semantic information. Therefore, pre-trained word vectors are used to initialize GCN, which is responsible for representing data information into low-dimensional vectors. GCN uses a stable two-layer GCN layer, with the output of each layer serving as the input data for the next layer and initializing its weights.
[0019] At the same time, attribute embedding continues to use pre-trained word vectors to generate attribute representations. Specifically, the CBOW continuous bag-of-words approach under Word2Vec is used to model attribute pre-trained word vectors, and the top 3,000 frequent attributes are selected. This avoids the high discreteness of attribute data caused by too few attributes. A two-layer GCN is used in conjunction with the high-speed gating mechanism HighwayGates for denoising.
[0020] A further technical solution is that in step 5, after noise optimization, the structure embedding is trained, and then the attributes are trained independently, and finally the parameters are The embedding vectors of structure and attributes are concatenated, and the concatenation formula is Obtain the overall structure;
[0021] Among them, H represents the final embedded representation after splicing; H se is the structural embedding representation; H attr represents attribute embedding representation; || represents the concatenation operation of vectors.
[0022] An embodiment of the present invention provides an entity alignment method based on noise optimization and attribute fusion. On the one hand, this method performs noise optimization processing on the labeled seed entity pairs according to the adversarial principle to obtain relatively pure seed entity pairs, in preparation for the next step of structural embedding training. Secondly, two layers of GCN based on the Highway Gates high-speed gating mechanism are used to embed the pre-trained word vector to obtain the entity representation. On the other hand, based on the contribution of attribute information to the research on entity alignment, combined with the optimization strategy of noise seed entity pairs, a new entity alignment model based on noise optimization and attribute fusion is constructed, which uses the pre-trained word vector of the attribute to obtain the embedded representation of the attribute in a similar structural embedding manner. This method uses the combination of noise optimization and attributes to not only reduce the complexity, but also improve the effect of entity alignment, which has promoted the entity alignment task and achieved excellent entity alignment effect on the basis of reducing experimental overhead. BRIEF DESCRIPTION OF THE DRAWINGS
[0023] Figure 1 A flowchart of an entity alignment method based on noise optimization and attribute fusion provided by an embodiment of the present invention;
[0024] Figure 2 A GCN propagation structure diagram of an entity alignment method based on noise optimization and attribute fusion provided by an embodiment of the present invention;
[0025] Figure 3 A global structural diagram of an entity alignment method based on noise optimization and attribute fusion provided by an embodiment of the present invention. DETAILED DESCRIPTION
[0026] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.
[0027] The specific implementation of the present invention is described in detail below with reference to specific embodiments.
[0028] like Figure 1-3 As shown, an entity alignment method based on noise optimization and attribute fusion provided by an embodiment of the present invention includes the following steps:
[0029] Step 1: Initialize the entity seed set with noise to obtain accurate noise entity pairs;
[0030] Step 2: Inject the noise-aware optimization generator to obtain more noise entity pairs around the relevant distribution of the true seed entity pairs;
[0031] Step 3: After the noise entity pairs are generated, they are fed into the noise-aware optimization discriminator to identify the true seed entity pairs and the noise entity pairs;
[0032] Step 4: The noise-aware optimized generator and the noise-aware optimized discriminator iterate each other to obtain a trust score and compare it with the threshold θ to obtain clean labeled data;
[0033] Step 5: The obtained real seed entity pairs are structurally embedded through GCN. The GCN embedding method is as follows: Figure 2 , and finally obtain the structural alignment model.
[0034] In the embodiment of the present invention, in step 1, it is not yet known which of the labeled seed entities are incorrect seed entity pairs. In order to enable the model to accurately identify noise entities and provide initialization data for the identification model to improve the probability of noise extraction, it is necessary to manually generate accurate noise entity pairs in advance. A certain proportion of seed entity pairs are extracted from the seed set containing the labels and randomly replaced. After the replacement is completed, they are merged with the remaining entity pairs to obtain accurate noise entity pairs.
[0035] In step 2, after initialization, the noise-aware optimization generator acquires noise entities around the distribution of true seed entity pairs to obtain more noise entity pairs. This acquisition process ensures that the noise entities are as close as possible to the true seed entity pairs to deceive the noise-optimized discriminator model. The probabilities of the entity pairs generated by the noise-aware optimization generator are based on negative samples, which are generated using Bernoulli negative sampling.
[0036] In step 4, the correct seed entity pair is determined by the threshold θ during the training model phase. When the classification function is greater than or equal to the threshold, the score function value is recorded as 1, that is, the classification function sigmoid (f τ (e i , e j ))≥θ, the score function Score(e i , e j ) is recorded as 1; that is, the current authentication entity is considered correct
[0037] Entity, on the contrary, when the classification function is less than the threshold, the score function value is recorded as 0, that is, the classification function is recorded as sigmoid (f τ (e i , e j ))<θ, score function Score(e i , e j ) is recorded as 0; the number of correct entity seeds is continuously expanded. The entity pairs generated by the noise-aware optimization generator are incorrect entity pairs, and the noise-aware optimization discriminator evolves into a binary classification problem to distinguish between positive and negative samples.
[0038] In step 5, pre-trained word vectors are used as the input for structural embedding. Compared to traditional structural information conversion and encoding, pre-trained word vectors contain rich semantic information. Therefore, pre-trained word vectors are used to initialize GCN. GCN is responsible for representing data information into low-dimensional vectors. GCN uses a stable two-layer GCN layer. The output of each layer is the input data of the next layer and initializes its weights.
[0039] At the same time, attribute embedding continues to use pre-trained word vectors to generate attribute representations. Specifically, the CBOW continuous bag-of-words approach under Word2Vec is used to model attribute pre-trained word vectors, and the top 3,000 frequent attributes are selected. This avoids the high discreteness of attribute data caused by a small number of attributes. A two-layer GCN is used in conjunction with the high-speed gating mechanism HighwayGates for denoising.
[0040] In step 5, after noise optimization, the structural embedding is trained, and then the attributes are trained independently, and finally the parameters are The embedding vectors of structure and attributes are concatenated, and the concatenation formula is Get the overall structure. Among them, H represents the final embedded representation after splicing; H se is the structural embedding representation; H attr represents attribute embedding representation; || represents the concatenation operation of vectors.
[0041] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.
Claims
1. An entity alignment method based on noise optimization and attribute fusion, characterized in that: The following steps are involved: Step 1: Initialize the entity seed set with noise to obtain accurate noise entity pairs; Step 2: Inject the noise-aware optimization generator to obtain more noise entity pairs around the relevant distribution of the true seed entity pairs; Step 3: After the noise entity pairs are generated, they are fed into the noise-aware optimization discriminator to identify the true seed entity pairs and the noise entity pairs; Step 4: The noise-aware optimized generator and the noise-aware optimized discriminator iterate each other to obtain a trust score and compare it with the threshold θ to obtain clean labeled data; Step 5: Use GCN to embed the obtained real seed entity pairs into structures and finally obtain the structural alignment model; In step 1, in the initial state, accurate noise entity pairs need to be artificially generated in advance; some seed entity pairs are extracted from the seed set containing labels and randomly replaced, and after the replacement is completed, they are merged with the remaining entity pairs to obtain accurate noise entity pairs; In step 2, after initialization is completed, in order to obtain more noise entity pairs, the noise-aware optimization generator needs to obtain noise entities around the relevant distribution of the real seed entity pairs. The noise-aware optimization generator generates entity pairs based on the probability of negative samples, and the generation of the negative samples adopts the Bernoulli negative sampling method; In step 5, the pre-trained word vector is used as the input part of the structure embedding, and the pre-trained word vector is initialized as the GCN. The GCN is responsible for representing the data information into a low-dimensional vector. The GCN adopts a stable two-layer GCN layer, and the output of each layer is the input data of the next layer, and its weight is initialized. At the same time, for attribute embedding, pre-trained word vectors are also used to generate attribute representations. Specifically, the CBOW continuous bag-of-words method under Word2Vec is used to model attribute pre-trained word vectors, and the top 3,000 frequent attributes are selected. A two-layer GCN is used in conjunction with the high-speed gating mechanism HighwayGates for denoising.
2. The entity alignment method based on noise optimization and attribute fusion according to claim 1 is characterized in that: In step 4, during the model training phase, the correct seed entity pair is determined by the threshold θ. When the classification function is greater than or equal to the threshold, the score function value is recorded as 1, that is, the classification function sigmiod (f τ (e i ,e j ))≥θ, the score function Score(e i , e j ) is recorded as 1; that is, the current identification entity is considered to be the correct entity. On the contrary, when the classification function is less than the threshold, the score function value is recorded as 0, that is, the classification function is recorded as sigmiod(f τ (e i , e j ))<θ, score function Score(e i , e j ) is recorded as 0; the number of correct entity seeds is continuously expanded, the entity pairs generated by the noise-aware optimization generator are incorrect entity pairs, and the noise-aware optimization discriminator evolves into a two-classification problem to distinguish positive and negative samples.
3. The entity alignment method based on noise optimization and attribute fusion according to claim 1, characterized in that: In step 5, after noise optimization, the structural embedding is trained, and then the attributes are trained independently, and finally the parameters are The embedding vectors of structure and attributes are concatenated, and the concatenation formula is Obtain the overall structure; Among them, H represents the final embedded representation after splicing; H se is the structural embedding representation; H attr represents attribute embedding representation; || represents the concatenation operation of vectors.