Relation Classification via Target-Concentrated Attention CNNs

Jizhao Zhu, Jianzhong Qiao, Xinxiao Dai, Xueqi Cheng

Abstract: Relation classification is a key natural language processing task that receives much attentions these years. The goal is to assign pre-defined relation labels to the nominal pairs marked in given sentences. It is obvious that different words in a sentence are differentially informative. Moreover, the importance of words is highly relation-dependent, i.e., the same word may be differentially important for different relations. To include sensitivity to this fact, we present a novel model, referred to as TCA-CNN, which takes the attention mechanism at the word level to pay different attention to individual words according to the semantic relation concentrated when constructing the representation of a sentence. Experimental results show that TCA-CNN achieves a comparable performance compared with the state-of-the-art models on the SemEval 2010 relation classification task.

External IDs:dblp:conf/iconip/ZhuQDC17