Definition Extraction with LSTM Recurrent Neural Networks

Abstract

Definition extraction is the task to identify definitional sentences automatically from unstructured text. The task can be used in the aspects of ontology generation, relation extraction and question answering. Previous methods use handcraft features generated from the dependency structure of a sentence. During this process, only part of the dependency structure is used to extract features, thus causing information loss. We model definition extraction as a supervised sequence classification task and propose a new way to automatically generate sentence features using a Long Short-Term Memory neural network model. Our method directly learns features from raw sentences and corresponding part-of-speech sequence, which makes full use of the whole sentence. We experiment on the Wikipedia benchmark dataset and obtain 91.2 % on $F_1$ score which outperforms the current state-of-the-art methods by 5.8 %. We also show the effectiveness of our method in dealing with other languages by testing on a Chinese dataset and obtaining 85.7 % on $F_1$ score.