Argument Mining for Medical Reviews

Abstract

Argument mining is the process of extracting opinions and reasons from dialectical text and
drawing conclusions to illuminate the author’s viewpoint concisely. Hence, argument mining
becomes highly useful in the medical domain, especially for pharmacists and analysts in
analysing the effects of drugs on people and their varying opinions on the effectiveness of the
drugs in question. In this paper, we propose a system that uses argument mining and machine
learning to extract supporting and attacking relationships between sentences from drug
reviews, in an effort to build an application that can provide deeper insight into people’s
opinions on various drugs. We identify argumentative content based on the presence of
discourse indicators, which then undergoes pre-processing and feature extraction to form a
meaningful representation of the text. We consider seven feature sets consisting of structural
features, TF-IDF scores for unigrams and bigrams and their combinations. The feature vectors
are given to a machine learning classifier for predicting support/attack relations between
sentence pairs. We evaluate three classification algorithms, namely support vector machine,
random forest classifier and AdaBoost classifier, using precision, recall, F1 scores and 10-fold
cross validation accuracy as evaluation parameters. The application can then give a detailed
analysis of the given medical review.