Sequence models and ranking methods for discourse parsing

• Authors: James Pustejovsky;Ben Wellner
• Venue: Sequence models and ranking methods for discourse parsing
• Year: 2009

Many important aspects of natural language reside beyond the level of a single sentence or clause, at the level of the discourse, including: reference relations such anaphora, notions of topic/focus and foreground/background information as well as rhetorical relations such as CAUSATION or MOTIVATION. This dissertation is concerned with data-driven, machine learning-based methods for the latter - the identification of rhetorical discourse relations between abstract objects, including events, states and propositions. Our focus is specifically on those relations based on the semantic content of their arguments as opposed to the intent of the writer. We formulate a dependency view of discourse in which the arguments of a rhetorical relation are lexical heads, rather than arbitrary segments of text. This avoids the difficult problem of identifying the most elementary segments of the discourse. The resulting discourse parsing problem involves the following steps: (1) identification of discourse cue phrases that signal a rhetorical relation (2) identification of the two arguments of a rhetorical relation signaled by a discourse cue phrase and (3) determination of the type of the rhetorical relation. To address the above problems, we apply a set of discriminative, statistical machine learning algorithms and explore the tradeoffs with various sets of features. We demonstrate how performance can be improved through learning architectures that allow for multiple co-dependent processing stages to be handled within a single model, rather than as a cascade of separate models. We capture additional dependencies with the novel application sequence-structured Conditional Random Fields to the problem of identifying discourse relations and their rhetorical types. The proposed Conditional Random Field model is more general than typically utilized in the literature, making use of non-factored feature functions to arrive at a conditional, sequential ranking model. Finally, we demonstrate the general applicability of our proposed discourse parsing model by applying it to the problem of syntactic dependency parsing, itself an important determinant for discourse parsing. This points towards a layered sequential (re-)ranking architecture for complex language processing tasks applicable beyond discourse parsing.