Improving Entity Linking in Chinese Domain by Sense Embedding Based on Graph Clustering

Entity linking refers to a string in text corresponding entities knowledge base through candidate entity generation and ranking. It is of great significance some NLP (natural language processing) tasks, such as question answering. Unlike English linking, Chinese requires more consideration due the lack spacing capitalization sequences ambiguity characters words, which evident certain scenarios. In domains, industry, generated are usually composed long strings heavily nested. addition, meanings words that make up industrial sometimes ambiguous. Their semantic space subspace general word embedding space, thus each needs get its exact meanings. Therefore, we propose two schemes achieve better linking. First, implement an n-gram based method increase recall rate reduce nesting noise. Then, enhance ranking mechanism by introducing sense embedding. Considering contradiction between vectors single domain, design model on graph clustering, adopts unsupervised approach for induction learns representation conjunction with context. We test quality our classical datasets demonstrate disambiguation ability confirm can learn entities’ fundamental laws domain performance experiments.