Combining Contextual Information Sources for Disambiguation in Parsing and Choice in Generation
PIs: Jonas Kuhn, Christian Rohrer (former)
Researchers: Martin Forst (former), Aoife Cahill (former), Sina Zarriess (former), Özlem Cetinoglu, Kyle Richardson
Project D2 is investigating the context-driven selection of LFG analyses from a set of candidates, either when parsing a string or when generating from an underlying input, such as an LFG f(unctional) structure. Contextual factors are modelled as features in a log-linear ranking model.
One of the major steps in the first phase was that we developed a statistical surface realisation ranking architecture based on the LFG generator. We showed that a linguistically informed feature model that incorporates information status outperforms previous models for ranking word order alternatives.
In the second phase, we have extended the scope of the generation architecture to choices that include other syntactic alternations (e.g. voice alternations) and referring expressions. On the parsing side, we have recently started to look at integrations of statistical dependency parsing and LFG-based syntax.