Research On Knowledge Reasoning Method Based On Knowledge Representation

Knowledge graph reasoning technology plays an important role in artificial intelligence related fields such as information retrieval,question answering systems and intelligent dialogue.Knowledge reasoning based on knowledge representation embeds entities and relationships into a continuous low-dimensional semantic space,effectively solves the problem of data sparseness,and efficiently calculates the semantic connection between entities and relationships,so that the performance of knowledge acquisition,fusion and reasoning significantly promote.The existing knowledge representation models generally only pay attention to the characteristics of the triple structure itself,and ignore the multi-modal information such as the visual information of the entity image and entity category information which play an important role in improving the performance of the knowledge representation learning model.In order to solve this problem,this paper proposes a multi-modal knowledge representation method MGAT model which combines entity image information and entity category information,combining the knowledge representation of multi-modal information with an improved GAT model.This model not only pays attention to the characteristics of the triple structure,but also takes into account the entity category information in the knowledge graph and the rich visual information outside the knowledge graph,and unifies the entity structure-based knowledge representation,image-based knowledge representation and category-based knowledge representation to carry out joint training to complete multi-modal knowledge graph representation.The main work is as follows:1.For entity images,this paper designs an image encoder to complete the extraction of entity image feature information and the conversion from image space to knowledge space,and uses the attention mechanism to carry out the entity's multi-image learning model to build the entity image-based representations;2.For entity category information,this paper uses the attention mechanism to model the semantic relationship between the entity category and the corresponding triple relationship,and construct the category-based knowledge representation;3.When training the model,this paper uses improved graph attention network to dig deeper entity characteristics,obtains stable knowledge expression,and uses it for subsequent related tasks.This paper compares and evaluates the model with the classic baseline model on the classic task link prediction and triple classification tasks of knowledge representation learning.The experimental results fully show that the rich information contained in multi-modal information can improve the expressive ability of the knowledge representation model,and it also shows that knowledge representation learning model fused with multi-modal information can make full use of this information to achieved better experimental results in actual knowledge learning tasks.