A Customized Attention-Based Long Short-Term Memory Network for Distant Supervised Relation Extraction

摘要

Distant supervision, a widely applied approach in the field of relationrnextraction can automatically generate large amounts of labeled trainingrncorpus with minimalmanual effort. However, the labeled training corpusrnmay havemany false-positive data, whichwould hurt the performance ofrnrelation extraction. Moreover, in traditional feature-based distant supervisedrnapproaches, extraction models adopt human design features withrnnatural language processing. It may also cause poor performance. To addressrnthese two shortcomings, we propose a customized attention-basedrnlong short-term memory network. Our approach adopts word-level attentionrnto achieve better data representation for relation extraction withoutrnmanually designed features to perform distant supervision instead ofrnfully supervised relation extraction, and it utilizes instance-level attentionrnto tackle the problem of false-positive data. Experimental resultsrndemonstrate that our proposed approach is effective and achieves betterrnperformance than traditional methods.