摘要
在自然语言处理中,实体指代项识别通常被看作是序列标注任务进行处理。而实体指代项又是由多个连续的序列标注子任务组成的。这些子任务包括切分识别、实体识别和指代项识别。传统的管道方法经常会导致每步间的错误向下传递。级联方式的联合模型会带来大量的标记。虚结点方法同时规避了上面两个方法的缺点。系统采用虚结点的在线联合模型,不仅减少了训练时间,性能也比普通的在线联合模型要好。
Entity mention recognition is always processed as sequence labelling task in natural language processing and the entity mention is composed of several sequence labelling subtasks.These subtasks are segmentation recognition,entity recognition and mention recognition.Traditional pipeline approaches often lead to error downward propagation between subtasks.Joint model with cascaded structures may introduce large amounts of label.The virtual node approach avoids the two problems above.My system,using online joint model with virtual node,not only reduces the training time,but is also proved to be better than the ordinary online joint models in performance.
出处
《计算机应用与软件》
CSCD
北大核心
2012年第4期192-194,219,共4页
Computer Applications and Software
关键词
实体指代识别
虚结点
级联
Entity mention recognition Virtual node Cascade