An empirical convolutional neural network approach for semantic relation classification

详细信息    查看全文

• 作者：Pengda Qin ^{qinpengda0406@gmail.com" class="auth_mail" title="E-mail the corresponding author}Author Vitae ; Weiran Xu ^{xuweiran@bupt.edu.cn" class="auth_mail" title="E-mail the corresponding author}Author Vitae ; Jun Guo ^{guojun@bupt.edu.cn" class="auth_mail" title="E-mail the corresponding author}Author Vitae
• 关键词：Relation classification ; Convolution neural network ; Dropout ; Data-driven
• 刊名：Neurocomputing
• 出版年：2016
• 出版时间：19 May 2016
• 年：2016
• 卷：190
• 期：Complete
• 页码：1-9
• 全文大小：835 K

文摘

In industry, relation classification plays a significant role in today׳s search engine. Up to now, the state-of-the-art systems have the problems of over-reliance on the quality of handcrafted features annotated by experts and linguistic knowledge derived from linguistic analysis modules, which is costly and leads to the issue of error propagation. Currently, with the data-driven approaches attracting wide attention, deep learning achieves impressive performance in semantic processing tasks without much effort on costly features. In this work, we deal with the relation classification task utilizing a convolutional neural network (CNN) approach to automatically control feature learning from raw sentences and minimize the application of external toolkits and resources. Our proposed method has several distinct features. First, we exploit a simple but rational way to specify which input tokens are the target nominals in the input sentence, instead of Position Feature that used in other neural network relation classification systems. Secondly, a most suitable dropout strategy is used to prevent units in the neural network from co-adapting too much, which significantly reduces over-fitting and improves the performance. Eventually, using only word embedding as input features is sufficient to achieve desirable performance. Our experiments on the SemEval-2010 Task-8 dataset show that our CNN architecture without using any additional extracted features significantly outperforms the state-of-the-art systems and achieves an F1-score of 84.8% only considering the context between the two target nominals.