Abstract
In sentence-level relation classification field, both recurrent neural networks (RNN) and conventional neural networks (CNN) have won tremendous success. These methods do not rely on NLP systems like named entity recognizers (NER). However either CNN or RNN has its advantages and disadvantages for relation classification. For example, CNN is good at capturing local feature, but RNN is good at capturing temporal features, particularly handling long-distance dependency between nominal pairs. This paper proposes BiLSTM-CNN model combining CNN and RNN, and compares it with CNN and RNN respectively. BiLSTM-CNN utilizes LSTM to extract series of higher level phrase representations, and then fed into CNN to do the relation classification. We conducted exhaustive research on two datasets: SemEval-2010 Task 8 (https://docs.google.com/View?docid=dfvxd49s_36c28v9pmw) dataset and KBP37 (https://github.com/zhangdongxu/kbp37) dataset. The result strongly indicates the BiLSTM-CNN has the best performance among models in the literature, particularly for long-span relations. And on KBP37 dataset, we achieve the state-of-the-art F1-score.
Published Version
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call