基于上下文信息和排序学习的实体链接方法

为解决英语命名实体链接问题,提出了一种基于上下文信息和排序学习的实体链接方法. 首先使用上下文信息对实体指称进行扩充,并在维基百科中检索候选实体列表;然后通过抽取实体指称与候选实体之间的各类特征,利用ListNet排序算法对候选实体列表进行排序,选出Top1的候选实体作为链接结果;最后对未找到候选的实体指称即NIL实体,通过实体聚类算法进行关联链接. 实验结果表明,该方法在KBP 2013实体链接数据集上的F值为0.660,比KBP 2013实体链接评测中所有参赛队伍的平均F值高0.092,比系统BUPTTeam2013的F值高0.162.

An Entity Linking Approach Based on Context Information and Learning to Rank

English entity linking tasks play an important role in construction of semantic network and big knowledge base. An entity linking method based on local information and learning to rank algorithm was proposed. Firstly, the context information is well used for expanding mentions' name and retrieving can-didate entities from Wikipedia. Secondly, kinds of features are extracted between mentions and candi-dates and also the ListNet algorithm was used to rank the candidate entities to choose the most related en-tity as the linked objects. Finally, the NIL entities was clustered by clustering method. The method a-chieved 0. 660 F value on KBP 2013 Entity Linking dataset, it performs 0. 092 better than the median F value of all participated teams in KBP 2013 entity linking task and also performs 0. 162 better than BUPTT eam 2013, which is the baseline comparison system in the experiment.