Research On Knowledge Reference Algorithm In Chinese Knowledge Graph

As an important means to improve the knowledge graph,knowledge inference plays a key role in the process of constructing the knowledge graph.There are some problems in the Chinese knowledge graph,such as sparse data,uneven quality and so on.At the same time,due to the characteristics of Chinese text,knowledge inference is easily affected by word segmentation errors and there are fewer available training data sets in the Chinese field,which has a certain impact on the training of Chinese knowledge inference model.Nowadays,most knowledge inference models are based on neural networks,distributed knowledge representation learning and logic rules.These models have obvious information loss in the process of extracting feature to construct feature vector,which results in uneven quality of the extracted feature vector and low interpretability of reference result.In order to improve the effect of model,this paper has made improvements from the aspect of encoding entity feature vector and the interpretability of inference.(1)In order to reduce the semantic loss caused by encoding entity description,a new entity description encoder is designed by deep neural network combined with translation model.First,use BiLSTM to obtain the complete context of the entity description,which avoids the semantic loss caused by DKRL using only some high-frequency words in the description,then focus on the local feature words in the description through the attention mechanism,and finally train with the translation model to obtain the entity description vector.(2)The knowledge inference model based on neural networks does not consider the rich semantics contained in the description and type information associated with entity.This paper uses the above entity description encoder and TKRL algorithm to fuse entity description and entity type information respectively,and then uses neural tensor networks to model the knowledge graph and complete knowledge reference,and enhances the model reference effect by fusing multisource information(3)Due to the low interpretability of knowledge representation learning model based on complex path in path vector representation.In this paper,logic rules and relation weighting are introduced to optimize the path representation.The Horn rule mined by AMIE+ algorithm which to help build path improved the interpretability of reference.(4)Put the model into engineering application,design and construct a knowledge inference system.In this paper,the corresponding comparative experiments are set up for the proposed improvement ideas.By combining entity description information and entity type information,the accuracy rate in the triple classification task is improved by 4.06%.The MRR and Hits@10 indicators in the link prediction task reached 72.43% and 81.93%,which illustrates the effectiveness of multi-source information fusion.Using rules to guide path representation learning,The Hits@10 indicator in the link prediction task reached 82.15%,which shows that the way of using rule-guided path representation not only improves the interpretability,but also significantly improves the inference effect.