Neural approaches to discourse coherence: modeling, evaluation and application

Discourse coherence is an important aspect of text quality that refers to the way different textual units relate to each other. In this thesis, I investigate neural approaches to modeling discourse coherence. I present a multi-task neural network where the main task is to predict a document-level coherence score and the secondary task is to learn word-level syntactic features. Additionally, I examine the effect of using contextualised word representations in single-task and multi-task setups. I evaluate my models on a synthetic dataset where incoherent documents are created by shuffling the sentence order in coherent original documents. The results show the efficacy of my multi-task learning approach, particularly when enhanced with contextualised embeddings, achieving new state-of-the-art results in ranking the coherent documents higher than the incoherent ones (96.9%). Furthermore, I apply my approach to the realistic domain of people’s everyday writing, such as emails and online posts, and further demonstrate its ability to capture various degrees of coherence. In order to further investigate the linguistic properties captured by coherence models, I create two datasets that exhibit syntactic and semantic alterations. Evaluating different models on these datasets reveals their ability to capture syntactic perturbations but their inadequacy to detect semantic changes. I find that semantic alterations are instead captured by models that first build sentence representations from averaged word embeddings, then apply a set of linear transformations over input sentence pairs. Finally, I present an application for coherence models in the pedagogical domain. I first demonstrate that state of-the-art neural approaches to automated essay scoring (AES) are not robust to adversarially created, grammatical, but incoherent sequences of sentences. Accordingly, I propose a framework for integrating and jointly training a coherence model with a state-of-the-art neural AES system in order to enhance its ability to detect such adversarial input. I show that this joint framework maintains a performance comparable to the state-of-the-art AES system in predicting a holistic essay score while significantly outperforming it in adversarial detection