A Multi-hop Reasoning Method and Device for Knowledge Graph Oriented to Attribute Fusion

Abstract

The present invention relates to a multi-hop reasoning method and device for a knowledge graph oriented to attribute fusion, belonging to the field of intelligent question answering, including: defining an attribute-oriented knowledge graph and multi-hop reasoning; the knowledge graph is used for structurally storing and representing entities of various objective facts and the attributes of the entities; multi-hop reasoning is used for exploring semantic paths in the knowledge graph by multi-hop logical reasoning, spanning multiple entities and the relationships between entities to solve problems; building an AMR-KG architecture for realizing the organic combination of the attributes of entities and multi-hop reasoning, and integrating attributes into the knowledge graph and the multi-hop reasoning process. The present invention improves the accuracy and efficiency of multi-hop reasoning, enhances the model's semantic understanding ability of complex relationships, provides a new reasoning framework for knowledge graphs rich in attributes, and provides new ideas for subsequent research.

Description

Technical Field

[0001] The present invention belongs to the field of intelligent question answering, and particularly relates to a multi-hop reasoning method and device for a knowledge graph oriented to attribute fusion. Background Art

[0002] Knowledge graph reasoning is one of the core tasks of knowledge graph applications, and its purpose is to derive new knowledge from existing knowledge. Reasoning tasks can be divided into single-hop reasoning and multi-hop reasoning. Single-hop reasoning mainly focuses on complementing missing entities or relationships, while multi-hop reasoning is more complex. It requires reasoning among multiple entities and relationships to answer complex query questions. For example, given a query "find all entities related to A and connected to C through B", multi-hop reasoning is needed to complete it.

[0003] With the continuous expansion of the scale of knowledge graphs and the increasing complexity of application scenarios, the importance of multi-hop reasoning has become more prominent. However, existing multi-hop reasoning methods face many challenges when dealing with large-scale knowledge graphs, such as low reasoning efficiency, insufficient support for sparse graphs, and difficulty in handling complex logical queries. In addition, rich attribute information (especially spatio-temporal information) in knowledge graphs is often ignored, and this attribute information is of great significance for understanding the relationships between entities.

[0004] In the research of knowledge graph multi-hop reasoning, there are mainly the following three categories of methods:

[0005] (1) Reasoning based on rule logic

[0006] This type of method performs logical reasoning using a pre-defined reasoning rule and a reasoning engine. For example, Datalog and Prolog are two typical reasoning tools based on rule logic. Datalog is a logic programming language that can perform reasoning through recursive rules, while Prolog is a more general logic programming language suitable for complex logical reasoning tasks. However, the disadvantages of this type of method are extremely high labor costs and difficulty in expanding to other knowledge graphs.

[0007] (2) Reasoning based on path search

[0008] This type of method starts from the starting node, constructs a search space in the knowledge graph, and moves forward along possible relationship paths to find the target node. For example, the A search algorithm is used for knowledge graph reasoning, and the path search is optimized through heuristic methods. In addition, reinforcement learning technology is also used to dynamically search for the optimal reasoning path. The advantage of this type of method is that it can flexibly handle various complex queries, but the disadvantages are large computational overhead, especially low efficiency in large-scale graphs and long-path searches. In addition, the performance of this type of method on sparse graphs is also limited.

[0009] (3) Embedding-based Reasoning

[0010] This type of method embeds entities and relationships into a low-dimensional vector space and performs reasoning through vector operations. For example, the TransE model represents entities and relationships as vectors and uses vector addition for reasoning. Query2box processes complex queries with uncertainty by designing hyper-rectangular boxes. In addition, some studies embed entities and relationships based on beta distributions and geometric models. The advantage of this type of method is strong scalability and the ability to adapt to large-scale knowledge graphs. However, a common drawback of existing methods is that they ignore the rich attribute information in the knowledge graph, especially spatio-temporal attributes, which are crucial for understanding the relationships between entities.

[0011] In terms of attribute embedding, existing studies have attempted to incorporate attribute information into knowledge graph embedding. For example, TransEA improves the reasoning performance of relationships between entities by combining discrete relationships and continuous numerical attributes through a joint embedding framework. AEKE proposes an entity attribute embedding framework for error perception to reduce incorrect reasoning. KGFA enhances the performance of recommendation systems through an attribute embedding mechanism. These studies provide ideas for using attribute information in knowledge graph reasoning, but have not systematically applied attribute information to multi-hop reasoning tasks.

[0012] Existing methods have the following defects: Existing methods ignore the rich attribute information in the knowledge graph, especially spatio-temporal attributes. The reasoning efficiency and accuracy are insufficient, especially in large-scale graphs and complex query scenarios. Ignoring attribute information: Most existing multi-hop reasoning methods for knowledge graphs do not fully utilize the rich attribute information (such as spatio-temporal attributes) in the knowledge graph, resulting in an incomplete characterization of entity semantics during reasoning and affecting reasoning accuracy. Insufficient reasoning efficiency and accuracy: Path search-based methods have high computational overhead and are inefficient, especially in large-scale graphs and complex queries; while embedding-based methods, although highly scalable, are limited in reasoning accuracy due to the lack of support from attribute information. Summary of the Invention

[0013] To solve the above technical problems, the present invention proposes a multi-hop reasoning method for knowledge graphs oriented to attribute fusion. The specific technical solution is as follows:

[0014] Step 1: Define an attribute-oriented knowledge graph and multi-hop reasoning; the knowledge graph is used to structurally store and represent entities of various objective facts and the attributes of the entities; multi-hop reasoning is used to explore semantic paths in the knowledge graph by performing multi-hop logical reasoning across multiple entities and the relationships between entities to solve problems;

[0015] Step 2: Build the AMR-KG architecture to achieve the organic combination of entity attributes and multi-hop reasoning, and integrate attributes into the knowledge graph and the multi-hop reasoning process.

[0016] The present invention also provides a knowledge graph multi-hop reasoning device for attribute fusion, including:

[0017] A definition module that defines a knowledge graph and multi-hop reasoning oriented to attributes; the knowledge graph is used to structurally store and represent entities of various objective facts and the attributes of the entities; multi-hop reasoning is used to explore the semantic paths in the knowledge graph through multi-hop logical reasoning, spanning multiple entities and the relationships between entities to solve problems;

[0018] An AMR-KG architecture that builds the AMR-KG architecture to achieve the organic combination of entity attributes and multi-hop reasoning, and integrate attributes into the knowledge graph and the multi-hop reasoning process.

[0019] The present invention also provides an electronic device, including: one or more processors; a memory for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the above-mentioned knowledge graph multi-hop reasoning method for attribute fusion.

[0020] The present invention also provides a computer-readable storage medium, on which executable instructions are stored, and when the instructions are executed by a processor, the processor implements the above-mentioned knowledge graph multi-hop reasoning method for attribute fusion.

[0021] The present invention has the following beneficial effects:

[0022] (1) Make full use of attribute information

[0023] Most of the existing technologies ignore the rich attribute information in the knowledge graph, especially spatio-temporal attributes, and this information is crucial for understanding entity relationships. The present invention integrates attribute information into entity embeddings through an entity attribute message passing network (EAMN) and a graph attention mechanism, enriching the semantic representation of entities, thereby significantly improving the accuracy of multi-hop reasoning.

[0024] (2) Provide a new reasoning framework

[0025] The existing technologies have deficiencies in processing knowledge graphs with rich attributes. The present invention proposes a multi-hop reasoning framework for attribute fusion (AMR-KG), systematically introducing attribute information into the multi-hop reasoning task, opening up new ideas and methods for reasoning in knowledge graphs with rich attributes, and having stronger adaptability and scalability. By introducing attribute information and attention mechanisms, the present invention can better capture the complex relationships between entities, especially performing well in knowledge graphs with rich attributes, significantly enhancing the model's ability to understand complex logical query problems. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 It is a structural diagram of a multi-hop reasoning method for a knowledge graph oriented to attribute fusion. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0027] In order to make the objectives, technical solutions and advantages of the present invention clearer and more understandable, 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 used to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other. To achieve the above objectives, the present invention adopts the following technical solutions.

[0028] As Figure 1 shown, it is a flowchart of a multi-hop reasoning method for a knowledge graph oriented to attribute fusion proposed by the present invention, including the following steps:

[0029] Step 1. Define an attribute-oriented knowledge graph and multi-hop reasoning; the knowledge graph is used to structurally store and represent entities of various objective facts and the attributes of the entities; multi-hop reasoning is used to explore the semantic paths in the knowledge graph by multi-hop logical reasoning, spanning multiple entities and the relationships between entities to solve problems;

[0030] Define one: Knowledge graph: The knowledge graph , where represents the entity set, represents the relationship set, represents the attribute set, represents the attribute value set, and the storage form of the knowledge graph consists of entity triples and attribute triples composed.

[0031] Define two: Multi-hop reasoning: Define the sample of multi-hop reasoning as a triple structure , which represents the answer to a question , the correct answer i.e., the positive sample and the wrong answer i.e., the negative sample set. In addition, the problem consists of first-order predicate logic, including conjunction (∧), disjunction (∨), existential quantification, ), and negation (¬). The problem q consists of multiple logical operations, and the goal is to find a target entity as the answer to the problem .

[0032] Step 2: Build an AMR-KG architecture for realizing the organic combination of entity attributes and multi-hop reasoning, and integrating attributes into the knowledge graph and the multi-hop reasoning process.

[0033] To introduce attributes to answer multi-hop reasoning questions, build an AMR-KG architecture as Figure 1 shown. The AMR-KG architecture includes a knowledge graph induction part, a sample making part, a question embedding part, entity embedding, and matching; among them, the knowledge graph induction part records the transitive relationship of multi-hop reasoning in the knowledge graph, as Figure 1 shown in (a); the sample making part makes positive and negative samples for the entities in the knowledge graph. The positive samples are obtained by positive sampling of the entities in the knowledge graph, and this process is called positive sampling. The negative samples are obtained by replacing the results of the entities in the knowledge graph, and this process is called negative sampling, as Figure 1 shown in (b); the question embedding part embeds the aggregated attributes related to the entity, as Figure 1 shown in (c), and finally calculates the KL divergence with the entity embedding for matching. Figure 1 Query the problem according to the structures considered in various ways in the knowledge graph in (a) of Figure 1 and instantiate the logical query problem according to these structures. In (b) of Figure 1 , the sampling part instantiates the logical query problem and the answer according to various structures as positive samples. At the same time, design a dynamic rejection sampling model for negative sample generation. The question embedding part includes an entity attribute message passing network and a question sequence network (QSN), Figure 1 shown in (c), the entity attribute message passing network is used to fuse the attributes related to the entity, Figure 1In (d), it is the entity embedding matching part. The KL divergence is calculated between the entity embedding vector obtained through sample learning and the question embedding vector output by the question sequence network to measure the matching degree. Finally, the distribution is generated using the entity embedding vector and the question embedding vector, and the KL divergence of the matching process is compared. It is expected that the distribution generated by the positive samples is closer, and the distribution difference between the negative samples is larger.

[0034] The dynamic rejection sampling model is used to generate more challenging negative samples to improve the discrimination ability of the model. The quality of the samples largely determines the training effect. Different from knowledge graph completion, the sampling process of multi-hop reasoning is much more complex. It includes generating questions by instantiating the query structure , performing KG traversal to find answers and negative answers . The loss calculation of the learning process of the dynamic rejection sampling model is shown as follows:

[0035] ,

[0036] where, represents the loss function, represents the parameters, represents the correct answer, represents the wrong answer, represents the parameters of the question embedding network, represents the parameters of the entity embedding, is the sigmoid function, represents the number of elements in the set. The loss consists of minimizing the distance between the question embedding and the answer and maximizing the distance between the question and the negative samples . γ is a hyperparameter used to adjust the distance scale. According to experience, positive samples can be obtained through graph traversal, but there need to be several times as many negative samples as positive samples for the network to learn. The learning process requires positive sampling and negative sampling.

[0037] To obtain more efficient negative samples in multi-hop reasoning, a dynamic rejection sampling model is proposed to replace the random entity negative sampling in the traditional method. Briefly, during the sampling process, the candidate distribution or acceptance criterion is continuously adjusted according to the current sampling results. Specifically, while recording the question content, the upper-level types corresponding to the entities mentioned in the question will also be recorded. The negative sampling of a question is divided into two parts. Under the upper-level types of the recorded question, entities of the same type as the recorded type are sampled as negative samples. The more entities of the same type, the larger the proportion of this part of the sampling points; the other part is completely random sampling. In this way, more entities of the same type that the model is prone to make mistakes on are obtained as negative samples for learning to improve the model accuracy.

[0038] The problem embedding part is implemented through an Entity Attribute Message Passing Network (EAMN). The problem is composed of multiple entities associated with each other. These entities will fuse attribute information for subsequent learning. Using a graph neural network and an attention mechanism, the attribute information is incorporated into the entities to enrich the semantic representation.

[0039] Using the message passing mechanism of attribute information, the process of problem embedding can be divided into two steps: message passing and state update, as shown in the following formula. Assume an entity whose current state entity embedding vector is , and the updated state is , the set of attribute values of , and the transmission of attribute information is the result of attribute state aggregation :

[0040] ,

[0041] ,

[0042] Among them, consists of a learnable matrix and an activation function, and is used to judge the proportion of state update. Since an entity is usually associated with different attributes under the semantics of different problems, the present invention designs an attention mechanism transmission framework for learning the influence of different attributes, and the specific calculation is as follows:

[0043] ,

[0044] ,

[0045] Among them, represents the activation function, represents the learnable parameter weight matrix. represents the attention coefficient calculation of entity to its attribute value , which is the result of softmax normalization of the attribute attention scores of each entity. and are both index values, indicating traversal, · represents the vector dot product calculation, and the final entity embedding vector is obtained through iteration. By introducing an attention-based attribute message passing network, the semantics contained in the attributes are utilized proportionally, introducing richer feature representations for subsequent multi-hop reasoning.

[0046] The entity that aggregates attribute information is holistically learned and embedded through the Question Sequence Network (QSN). After learning through the GRU and attention mechanism, the Question Sequence Network (QSN) matches the result of entity embedding with the KL divergence.

[0047] The Question Sequence Network embeds the entity attribute fusion result after fusing attributes, that is, the entity embedding vector, to generate a question embedding vector representing the question, including:

[0048] Unfolding the question into a triple question sequence:

[0049] ,

[0050] Among them, represents the length of the question sequence; using the encoding of GRU to capture sequence dependencies, the final length is The question embedding vector of the question sequence is represented as:

[0051] ,

[0052] Among them, is the output of the reset gate, σ is the sigmoid activation function, , , are trainable parameters, is the th entity 's entity embedding vector, is the hidden state of the previous question embedding vector in the question sequence, is the state of the hidden layer, , , are trainable parameters, ⊙ represents element-wise multiplication, is the hyperbolic tangent activation function, is the output of the update gate, , , are trainable parameters.

[0053] The calculation of the KL divergence is specifically as follows: First, represent the entity embedding vector and the question embedding vector as probability distributions, and convert the embedding vector into a probability distribution through the softmax function;

[0054] and are respectively the probability distribution representations of the entity embedding vector and the question embedding vector , then:

[0055] ,

[0056] ,

[0057] Among them, and are the i-th components of the entity embedding and question embedding vectors respectively; and are the j-th components of the entity embedding and question embedding vectors respectively;

[0058] Calculate and The calculation formula for the KL divergence of is:

[0059] .

[0060] Then calculate the KL divergence between them, which can measure the similarity between the entity and the question. The smaller the KL divergence, the more relevant the entity and the question are. This process involves the representation of probability distributions, the calculation of KL divergence, and the matching of entities and questions.

[0061] The present invention also provides a knowledge graph multi-hop reasoning device for attribute fusion, including:

[0062] A knowledge graph multi-hop reasoning device for attribute fusion, including:

[0063] A definition module that defines a knowledge graph for attributes and multi-hop reasoning; the knowledge graph is used for structured storage and representation of entities of various objective facts and the attributes of the entities; multi-hop reasoning is used to explore semantic paths in the knowledge graph through multi-hop logical reasoning, spanning multiple entities and the relationships between entities to solve problems;

[0064] An AMR-KG architecture that builds an AMR-KG architecture for realizing the organic combination of entity attributes and multi-hop reasoning, integrating attributes into the knowledge graph and the multi-hop reasoning process.

[0065] The present invention also provides an electronic device, including: one or more processors; a memory for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the above-mentioned knowledge graph multi-hop reasoning method for attribute fusion.

[0066] The present invention also provides a computer-readable storage medium, on which executable instructions are stored, and when the instructions are executed by a processor, the processor implements the above-mentioned knowledge graph multi-hop reasoning method for attribute fusion.

Claims

1. A knowledge graph multi-hop reasoning method for attribute fusion, characterized in that: The following steps are involved: Step 1: Define attribute-oriented knowledge graph and multi-hop reasoning; knowledge graph is used to structuredly store and represent various objective fact entities and entity attributes; multi-hop reasoning is used to explore semantic paths in knowledge graph through multi-hop logical reasoning, across multiple entities and relationships between entities, to solve problems; Step 2: Build the AMR-KG architecture to realize the organic combination of entity attributes and multi-hop reasoning, and integrate attributes into the knowledge graph and multi-hop reasoning process; The AMR-KG architecture includes knowledge graph induction, sample production, question embedding, entity embedding, and matching; The question embedding part includes an entity attribute message passing network and a question sequence network. The entity attribute message passing network is used to fuse entity-related attributes to obtain an entity attribute fusion result, namely, an entity embedding vector. The question sequence network embeds the entity attribute fusion result, namely, the entity embedding vector, to generate a question embedding vector representing the question. Problems in entity attribute message passing networks By multiple entities The multiple entities are associated Subsequent learning is performed by integrating attributes; attributes are transmitted using a message passing mechanism, and the problem embedding process is divided into two steps: message passing and state updating, as shown in the following formula: The entity embedding vector of the current state is , the entity embedding vector of the updated state is ,entity The attribute value set is , the attribute passed is the result of attribute state aggregation : ， ， in, Composed of a learnable matrix and activation function, the attention mechanism transfer framework Used to learn the impact of different attributes, the specific calculation is as follows: ， ， in, Represents all relevant attribute values ​​for entity e The weighted sum of represents the activation function, represents the learnable parameter weight matrix, Representing Entity The attribute value of Calculation of attention coefficient, and are all index values, indicating traversal, · represents vector dot product calculation, and the entity finally obtained after iteration The entity embedding vector is ; The question sequence network embeds the entity attribute fusion result after the fusion of attributes, that is, the entity embedding vector, to generate a question embedding vector representing the question, including: The problem Expands into a sequence of triple questions: ， in, The length of the problem sequence represented by GRU is used to capture the sequence dependency, and the final length is The question embedding vector of the question sequence is represented as: ， in, is the output of the reset gate, σ is the sigmoid activation function, , , is a trainable parameter, It is Entities The entity embedding vector of is the hidden state of the embedding vector of the previous question in the question sequence, is the state of the hidden layer, , , is a trainable parameter, ⊙ represents element-wise multiplication, is the hyperbolic tangent activation function, is the output of the update gate, , , is a trainable parameter.

2. According to claim 1, a knowledge graph multi-hop reasoning method for attribute fusion is characterized in that: Defining the Knowledge Graph ,in Represents a collection of entities, Represents a set of relations, Represents a collection of attributes. Represents a set of attribute values. The knowledge graph is stored in the form of entity triples. and attribute triples composition.

3. The knowledge graph multi-hop reasoning method for attribute fusion according to claim 2 is characterized in that: Define the sample of multi-hop reasoning as a triple structure , which represents a problem The answer, the correct answer Positive Sample and wrong answers Negative samples ;question It consists of several first-order predicate logic functions, including intersection, union, existence, and negation; The goal is to find a target entity As a question The answer.

4. The knowledge graph multi-hop reasoning method for attribute fusion according to claim 3 is characterized in that: in, The knowledge graph induction part records the entities in the knowledge graph and the relationships between entities; the sample production part produces positive and negative samples for the entities in the knowledge graph, where the positive samples are obtained by positive sampling of the entities in the knowledge graph, and the negative samples are obtained by replacing the entities in the knowledge graph; the question embedding part aggregates and embeds the relevant attributes of the entity; the entity embedding part calculates the KL divergence with the question embedding part and matches them.

5. The knowledge graph multi-hop reasoning method for attribute fusion according to claim 4 is characterized in that: A dynamic rejection sampling model is used for negative sample generation.

6. The knowledge graph multi-hop reasoning method for attribute fusion according to claim 5 is characterized in that: The loss calculation of the learning process of the dynamic rejection sampling model is as follows: ， in, represents the loss function, Represents the parameters, Represents the correct answer, Indicates a wrong answer. Represents the parameters of the question embedding part, Parameters representing the embedding part of the entity, is the sigmoid function, It represents the number of elements in the set, and γ is a hyperparameter used to adjust the distance scale.

7. The knowledge graph multi-hop reasoning method for attribute fusion according to claim 6 is characterized in that: The calculation of KL divergence is as follows: first, the entity embedding vector and the question embedding vector are represented as probability distributions by converting the embedding vectors into probability distributions through the softmax function; and are entity embedding vectors and the question embedding vector The probability distribution of is expressed as: ， ， in, and are the i-th component of the entity embedding and question embedding vectors respectively; and are the jth components of the entity embedding and question embedding vectors respectively; calculate and The calculation formula of KL divergence is: 。 8. A knowledge graph multi-hop reasoning device for attribute fusion, characterized in that: include: Define modules, define attribute-oriented knowledge graphs and multi-hop reasoning; Knowledge graphs are used to structure the storage and representation of various objective fact entities and their attributes; Multi-hop reasoning is used to explore semantic paths in the knowledge graph through multi-hop logical reasoning across multiple entities and relationships between entities to solve problems; AMR-KG architecture, build the AMR-KG architecture to realize the organic combination of entity attributes and multi-hop reasoning, and integrate attributes into the knowledge graph and multi-hop reasoning process; The AMR-KG architecture includes knowledge graph induction, sample production, question embedding, entity embedding, and matching; The question embedding part includes an entity attribute message passing network and a question sequence network. The entity attribute message passing network is used to fuse entity-related attributes to obtain an entity attribute fusion result, namely, an entity embedding vector. The question sequence network embeds the entity attribute fusion result, namely, the entity embedding vector, to generate a question embedding vector representing the question. Problems in entity attribute message passing networks By multiple entities The multiple entities are associated Subsequent learning is performed by integrating attributes; attributes are transmitted using the message passing mechanism, and the problem embedding process is divided into two steps: message passing and state updating, as shown in the following formula: The entity embedding vector of the current state is , the entity embedding vector of the updated state is ,entity The attribute value set is , the attribute passed is the result of attribute state aggregation : ， ， in, Composed of a learnable matrix and activation function, the attention mechanism transfer framework Used to learn the impact of different attributes, the specific calculation is as follows: ， ， in, Represents all relevant attribute values ​​for entity e The weighted sum of represents the activation function, represents the learnable parameter weight matrix, Representing Entity The attribute value of Calculation of attention coefficient, and are all index values, indicating traversal, · represents vector dot product calculation, and the entity finally obtained after iteration The entity embedding vector is ; The question sequence network embeds the entity attribute fusion result after the fusion of attributes, that is, the entity embedding vector, to generate a question embedding vector representing the question, including: The problem Expands into a sequence of triple questions: ， in, The length of the problem sequence represented by GRU is used to capture the sequence dependency, and the final length is The question embedding vector of the question sequence is represented as: ， in, is the output of the reset gate, σ is the sigmoid activation function, , , is a trainable parameter, It is Entities The entity embedding vector of is the hidden state of the embedding vector of the previous question in the question sequence, is the state of the hidden layer, , , is a trainable parameter, ⊙ represents element-wise multiplication, is the hyperbolic tangent activation function, is the output of the update gate, , , is a trainable parameter.

9. An electronic device, characterized in that: include: one or more processors; A memory for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the knowledge graph multi-hop reasoning method for attribute fusion as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that: Executable instructions are stored thereon, which, when executed by a processor, enable the processor to implement the knowledge graph multi-hop reasoning method for attribute fusion as described in any one of claims 1 to 7.

Images