TGCEL: A Chinese entity linking method based on topic relation graph

Abstract

Entity linking has an important basic research value for Natural Language Processing, the task of which is to link different entity mentions in the given text with their referent entities in a knowledge base. And it is widely used in such fields as expanding knowledge base, Q&A system, machine translation. We propose a Chinese collective entity linking algorithm based on the extracted topic features. We construct the topic relation graph of ambiguous entities in the same text, extract the topic characteristics from the multiple topic models, calculate the topic relevance, and select the topic subgraph with maximum score to reason and realize the batch linking. We experiment with both the news test corpus and the microblog test corpus, compare the performance of the adopted topic model, and analyze their applicable scene. When compared with the traditional algorithm, the maximum performance of our algorithm is improved by about 9% in microblog corpus and over 15% in news corpus, which indicates that our algorithm is potentially effective.