In Rhetorical Structure Theory, discourse units participate in asymmetric relationships, with one element acting as the nucleus and the other as the satellite. In the resulting tree-like nuclearity structure, the importance of each discourse unit can be measured by the number of relations in which it acts as the nucleus or as the satellite. Existing approaches to automatically parsing such structures suffer from two problems: they employ local inference techniques that do not capture document-level structural regularities, and they rely on annotated training data, which is expensive to obtain at the discourse level. We investigate the SampleRank structure learning algorithm as a potential solution to both problems. SampleRank allows us to incorporate arbitrary document-level features in a global stochastic inference algorithm. Furthermore, it enables the training of a joint model of discourse structure and summarization, which can be learned from document-level summaries alone, without discourse-level supervision. We obtain mixed results in the fully supervised case, and negative results for the joint model of discourse structure and summarization.
展开▼
