Early studies on discourse rhetorical structure parsing mainly adopt bottom-up approaches, limiting the parsing process to local information. Although current top-down parsers can better capture global information and have achieved particular success, the importance of local and global information at various levels of discourse parsing is different. This paper argues that combining local and global information for discourse parsing is more sensible. To prove this, we introduce a top-down discourse parser with bidirectional representation learning capabilities. Existing corpora on Rhetorical Structure Theory (RST) are known to be much limited in size, which makes discourse parsing very challenging. To alleviate this problem, we leverage some boundary features and a data augmentation strategy to tap the potential of our parser. We use two methods for evaluation, and the experiments on the RST-DT corpus show that our parser can primarily improve the performance due to the effective combination of local and global information. The boundary features and the data augmentation strategy also play a role. Based on gold standard elementary discourse units (EDUs), our parser significantly advances the baseline systems in nuclearity detection, with the results on the other three indicators (span, relation, and full) being competitive. Based on automatically segmented EDUs, our parser still outperforms previous state-of-the-art work.

Document-level machine translation (MT) remains challenging due to its difficulty in efficiently using document-level global context for translation. In this paper, we propose a hierarchical model to learn the global context for document-level neural machine translation (NMT). This is done through a sentence encoder to capture intra-sentence dependencies and a document encoder to model document-level inter-sentence consistency and coherence. With this hierarchical architecture, we feedback the extracted document-level global context to each word in a top-down fashion to distinguish different translations of a word according to its specific surrounding context. Notably, we explore the effect of three popular attention functions during the information backward-distribution phase to take a deep look into the global context information distribution of our model. In addition, since large-scale in-domain document-level parallel corpora are usually unavailable, we use a two-step training strategy to take advantage of a large-scale corpus with out-of-domain parallel sentence pairs and a small-scale corpus with in-domain parallel document pairs to achieve the domain adaptability. Experimental results of our model on Chinese-English and English-German corpora significantly improve the Transformer baseline by 4.5 BLEU points on average which demonstrates the effectiveness of our proposed hierarchical model in document-level NMT.