DeepCause: Hypothesis Extraction from Information Systems Papers with Deep Learning for Theory Ontology Learning

Abstract

This paper applies different deep learning architectures for sequence labelling to extract causes, effects, moderators, and mediators from hypotheses of information systems papers for theory ontology learning. We compared a variety of recurrent neural networks (RNN) architectures, like long short-term memory (LSTM), bidirectional LSTM (BiLSTM), simple RNNs, and gated recurrent units (GRU). We analyzed GloVe word embedding, character level vector representation of words, and part-of-speech (POS) tags. Furthermore, we evaluated various hyperparameters and architectures to achieve the highest performance scores. The prototype was evaluated on hypotheses from the AIS basket of eight. The F1 result for the sequence labelling task of causal variables on a chunk level was 80%, with a precision of 80% and a recall of 80%.