Most delphin-related work in Informatics at Edinburgh occurs within the context of the ROSIE project, standing for "RObust Semantic Interpretation". Ths project is funded by the Edinburgh Stanford Link. The overall aim is to provide robust, discourse-situated semantic interpretations through a combination of precise grammars and probabilistic techniques. Particular objectives include:

• Combining precise hand-crafted and probabilistic NLP models to obtain robust and deep discourse interpretations
• Building rich probabilistic NLP models which incorporate discourse- and semantic-conditioning
• Developing active learning methods for semi-automatic corpus annotation
• Investigating statistical methods well-suited to the sparse and noisy distributions prevalent in NLP datasets.

We have recently developed a head-driven probabilistic discourse parser, which assigns dialogues from the Verbmobil domain with its rhetorical structure. We have also investigated the utility of active learning in preparing treebanks for parse selection tasks. We are currently augmenting the rhetorical structure with its semantic consequences, resolving temporal anaphora and identifying communicative goals. This additional information will be added to the labelled corpus, in preparation for training a probablistic model of discourse interpretation.

We have also explored the formal framework of MRS and RMRS, in particular supplying a model theory and proving nice properties about semantic composition within these frameworks.

In the near future, we plan to work on creating semantic-based resources for languages that lack deep grammars, with the aid of parallel corpora.