Trigger word mining for relation extraction based on activation force
Weiran Xu, Chunyun Zhang · International Journal of Communication Systems · 2014
Summary In this paper, relation extraction is characterized as structured feature learning, and activation force (AF) is employed to extract and construct structured features. Trigger word is a low‐level feature, and it is very crucial in relation extraction. We define the trigger word as a word that is most likely to form a structure corresponding to a special relation. To extract trigger words, firstly, posteriori probability weighted frequency AF model, in which AF is regarded as posteriori probability weighted frequency, is presented. Secondly, a generative AF is introduced. Then, a higher level structured feature, named trigger‐word dependency pair (TWDP), is extracted by a reduced AF model. Based on the trigger words and TWDPs, the most advanced patterns, the shortest dependency paths (SDPs), are optimized. The evaluation corpus of Knowledge Base Population in Text Analysis Conference is adopted to test our methods. Experiments show that 87.30% of Stanford manual trigger words appearing in the training sentences can be found by G‐AF. The experimental results also verified that the modified SDP patterns are superior to original SDP patterns. Copyright © 2014 John Wiley & Sons, Ltd.