Advances on the automatic identification of signaled discourse relations

As a branch of computational linguistics, discourse analysis is concerned with uncovering and understanding organizational structures of texts. The Penn Discourse Treebank (PDTB), a collection of annotated Wall Street Journal articles, has been designed for the analysis of individual discourse relations connecting two text segments. The segments are designated as the relation’s arguments; the meaning between these arguments is its sense. This makes the approach more theory-neutral compared to other discourse frameworks that aim to build complex structures such as trees or graphs. Recognizing these discourse relations via shallow discourse parsing provides crucial information useful in numerous tasks, including machine translation, summarization, and simplification. This dissertation’s primary focus is on signaled discourse relations, in particular their identification and statistical applications. The PDTB distinguishes two signal variants: On the one hand, explicits are characterized by the use of well-studied discourse connectives (for example, and, however, due to). On the other hand, alternative lexicalized relations (AltLex) are realized by more complex phrases (for example, one particular reason for this), which are considered particularly challenging to automatically identify due to their rich variability and the limited data available. Because of recent advancements in deep learning, this thesis examines to what extent neural-based models can be utilized to recognize signaled discourse relations, in particular in order to identify discourse signals, label their arguments, and uncover their underlying sense. A novel framework is developed for designing shallow discourse parsing pipelines in a modular fashion. This framework incorporated methods presented in this thesis. It extends them by introducing contextualized word embeddings, clearly outperforming existing approaches on explicit relation recognition (both on PDTB and out-of-domain data). Furthermore, this thesis studies two challenging problems: First, semi-supervised learning methods are examined to overcome the commonly encountered data sparsity problem in discourse parsing. Two experiments concerning the recognition of alternative lexicalizations and the extraction of explicit arguments demonstrate positive trends in the models’ learning behavior as we improved on the baselines. Second, this thesis analyzes the rhetorical distinctions across domains and how to utilize them to differentiate documents of different domains. For this purpose, models developed in the first part of this thesis are applied across collected domain data such as argumentative essays, political speeches, and scientific abstracts. An analysis of the extracted signals uncovers varying usage of connectives and empirically demonstrates specific recurring sense patterns. Additionally, novel signal variants are recognized across domains that have not been previously observed in the PDTB data. Finally, the observed sense patterns are utilized to successfully distinguish documents across domains.