Research On Knowledge Graph Completion Based On Dense Feature Model

Knowledge base is a very important component of artificial intelligence,knowl-edge graph is the modern form of knowledge base and production of internet times.Artificial intelligence of new generation must be the combination of symbolism and connectionism and enhance to each other.Knowledge graph has been becoming very important resource to support the artificial intelligence system running with high performace,especially with more intelligence.However,although the scale of the knowledge graph is large,they are still incomplete.This disadvantage prevents them from widely application or even leads to fail.As a result,the problem of knowledge graph completion comes to forth.Recent years,with the development of represen-tation learning,there are fruitful results in this research line.Unfortunately,current knowledge grpah completion models designing seem encountering bottleneck that the complexity of new proposed models have been increased greatly while the per-formace does not promote as much as expected.This research mainly focus on the topic of improving the performace of the models at a low cost.On the basis of expor-ing the embedding-based models for knowledge graph completion,a comprehensive recogonition of the regular on these model is formed,and desstrutured them from the perspective of information thoery.A new paradigm is formulated for evaluating and designing this kind of models.In summary,there are five contributions in this paper:First,extracting multi-hop relation features by Hamming Distance.In this work,the model HRESCAL is propsod to capture the arbitrary length of relation features to increase the features for the succeding factoring by introducing Hamming Distance according to the mathematical character of adjacency matrix,and inreasing the den-sity for feature extraction.Second,increasing the freedom of model to improve the expressiveness of it.Since the translation equation is an approxmate one,increasing the freedom can make it becoming into an equal one and proposed a model Ros E.This enhands the expres-siveness of the model and can capture more feature to support better performace.Further research shows that the translation-based models transformed from transla-tion equation non-linearly still is optimistic.Third,coupling increasing the diversity of features extracting.Making fully use of translation-based and factorization-based models with different capacity to cap-ture distinctive features,and thus it can extracting more dense features.A model Trans RESCAL is proposed to improve the performance of the model.As a bypro-duction,the experimental results shows that the head and tail entity embeddings are determined artificially in factorization-based models.Fouth,modeling the essential features explicitly for a dataset.This model QLog-ic E is proposed by combining the translation-based and quantum logic with the in-tention of overcoming its low performance on the dataset WN18.Expectedly,the performace outperforms the state-of-the-art with large margin,but the space and time are kept in a low level,even breaking through the lower-bound pointed out in the work of quantum embedding.Fifth,deconstruction the expressiveness of the model with information thoery.The model QLogic E breaking though the lower-bound inspires the further exploring the KGC and propose the framework of DFM to characterize the expressiveness of embedding-based models.It can quantify the key factors in expressiveness and de-termine the criteria and limit value,which is beneficial for evaluating and designing the embedding-based models.