Deep learning for uncertainty detection
This project will tackle the problem of detecting uncertainty in texts, sometimes also referred to as hedge detection or speculation detection. The task is typically broken down into two subtasks; first identifying so-called cues, i.e., that function to indicate uncertainty (e.g., appear, indicate), and secondly identifying the scope, i.e., the span of text within the sentence affected by the uncertainty. In this project we want to assess the usefulness of applying so-called deep learning for this, i.e., we will train a sequence classifier based on a neural network architecture with several hidden layers of neurons (more specifically variants of models known as LSTM or GRU). For training and testing will rely on the BioScope corpus, consisting of articles and journals from the biomedical domain, manually annotated with information about speculation (and negation). For implementing the neural classifier we will be using Keras, a python library that can be used as a front-end for the deep learning platforms Theano or TensorFlow. This project is suited for students with a mathematical inclination.