An Efficient Hybrid Bi-LSTM Attention Model for Claims Extraction from Research Articles through Deep Learning

Abstract

The research literature is growing rapidly. A research article contains massive amounts of textual information. Claims are the most significant information in a research article that needs to be retrieved to understand the gist of the research work. A research article contains a number of claims in different sections (abstract, introduction, results, and discussion) of an article. A literature review shows that a few studies of claim extraction have been conducted and they are limited to extracting claims from the abstract section of the article only. In existing studies claims are classified either on the basis of the keywords or all the words. In existing works semantics and context are ignored, and Bag of Words (BoW) representation is used. Deep learning architectures such as Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) have the potential to produce better results through the use of deep learning. Attention mechanism and Bidirectional Long Short-Term Memory (Bi-LSTM) have been used for multiple tasks in Natural Language Processing (NLP) and give effective results. In this work, we propose a hybrid Bi-LSTM attention model. Bi-LSTM model captures the long-term sequences of words and the attention mechanism highlights the important words or keywords in text. A number of experiments have been performed on research articles claim and standard IBM datasets. We verified our proposed model on two datasets for claim/non-claim classification and our model gives 96.7% and 95% on both datasets. The results show that our proposed model through deep learning outperforms existing models.