Research On Entity Linking And Expansion Based On Enhanced Semantics And Trust Scores

Cognitive computing with natural language understanding as its core plays an important role in the field of artificial intelligence and has become an important consideration in the national and enterprise's artificial intelligence strategies.As there are a lot of heterogeneous text information on the Internet,how to better process and understand text from semantic level and then provide support for upper-layer applications in natural language processing and artificial intelligence such as text classification,intelligent QA and search recommendation is a challenging job.Entity linking aims to understand text semantics and the object these semantics describes and link entity mention to entity base,which can bridge the gap between natural language and machine understanding.By processing new entity found in entity linking and add them to entity base so that the timeliness and completeness of the entity base can be satisfied.So it is of great value to research entity linking and expansion.In order to solve the problem exists in entity linking and expansion,this paper studies the entity linking technology based on enhanced semantics and trust which improves the model's ability to represent semantics and linking performance.Besides,this paper encodes the vector representation of non-chained objects through the deep learning network and provides supports for entity expansion.Specifically,the main research contents and results of this paper includes three parts:(1)An enhanced entity representation which combines content and semantic relevance is proposed and custom context semantic vector training process is also implemented in this paper.This representation can make full use of entity' s text information,anchor links and entity co-occurrence relationships and avoids the lack of entity information when we rely solely on entity text content or entity references.Experiments conducted on entity relatedness shows that our model improves entity representation quality.(2)An entity-linking algorithm based on trust is proposed.By constructing disambiguation graph in each iteration,the background information of already disambiguated entity mentions are utilized in each iteration.Then trust score of each candidate nodes are evaluated and calculated and the performance on MicroP and MacroP is improved.(3)A context model of non-chained objects is proposed.The context words and entities are encoded by LSTM and multi-layer perceptron as the vector representation of non-chained objects.This vector representation is in the same vector space as the entity semantic model proposed in part 1 and the performance of non-chained obj ects representation and clustering are improved.