Structure features need complicated pre-processing, and are probably domain-dependent. To reduce time cost of pre-processing, we propose a novel neural network architecture which is a bi-directional long-short-term-me...Structure features need complicated pre-processing, and are probably domain-dependent. To reduce time cost of pre-processing, we propose a novel neural network architecture which is a bi-directional long-short-term-memory recurrent-neural-network（Bi-LSTM-RNN） model based on low-cost sequence features such as words and part-of-speech（POS） tags, to classify the relation of two entities. First, this model performs bi-directional recurrent computation along the tokens of sentences. Then, the sequence is divided into five parts and standard pooling functions are applied over the token representations of each part. Finally, the token representations are concatenated and fed into a softmax layer for relation classification. We evaluate our model on two standard benchmark datasets in different domains, namely Sem Eval-2010 Task 8 and Bio NLP-ST 2016 Task BB3. In Sem Eval-2010 Task 8, the performance of our model matches those of the state-of-the-art models, achieving 83.0% in F1. In Bio NLP-ST 2016 Task BB3, our model obtains F1 51.3% which is comparable with that of the best system. Moreover, we find that the context between two target entities plays an important role in relation classification and it can be a replacement of the shortest dependency path.展开更多
基金Supported by the China Postdoctoral Science Foundation(2014T70722)the Humanities and Social Science Foundation of Ministry of Education of China(16YJCZH004)
文摘Structure features need complicated pre-processing, and are probably domain-dependent. To reduce time cost of pre-processing, we propose a novel neural network architecture which is a bi-directional long-short-term-memory recurrent-neural-network（Bi-LSTM-RNN） model based on low-cost sequence features such as words and part-of-speech（POS） tags, to classify the relation of two entities. First, this model performs bi-directional recurrent computation along the tokens of sentences. Then, the sequence is divided into five parts and standard pooling functions are applied over the token representations of each part. Finally, the token representations are concatenated and fed into a softmax layer for relation classification. We evaluate our model on two standard benchmark datasets in different domains, namely Sem Eval-2010 Task 8 and Bio NLP-ST 2016 Task BB3. In Sem Eval-2010 Task 8, the performance of our model matches those of the state-of-the-art models, achieving 83.0% in F1. In Bio NLP-ST 2016 Task BB3, our model obtains F1 51.3% which is comparable with that of the best system. Moreover, we find that the context between two target entities plays an important role in relation classification and it can be a replacement of the shortest dependency path.
