23. 10. 2006
Keith Hall (Center for Language and Speech Processing, Johns Hopkins
University, USA)
Multilingual Dependency Parsing and Applications

Slidy audio

Abstract: Dependency parsing has recently come to the forefront of
interest in the statistical parsing community, culminating in the 2006
CoNLL shared task on multilingual dependency parsing. Many of the
competing teams made use of the Maximum Spanning Tree (MST) approach
pioneered by McDonald and Ribarov (McDonald et al. '05).
A disadvantage of the MST approach is that it requires structural scores
to be derived from parent-child links. This constrains the parsing models
to be based on very local structure; disallowing the explicit modeling of
subcategorization and valency as well as far simpler constraints (compound
adjectives, etc.).
In this talk, I present a two-stage dependency parser which combines a
K-best MST algorithm with a reranker. The advantage of such an approach is
that the model used by the reranker includes features defined over entire
tree structures. I present empirical results showing that “good” parses
appear in the first 50 hypotheses generated by the heavily constrained MST
models. Furthermore, I present reranking results that are competitive with
the state-of-the-art parsers. Results are presented for a subset of the
CoNLL competition languages as well as English.
Finally, I will introduce a framework for the application of dependency
parsing to tasks such as Speech Reconstruction and parsing of
resource-poor languages.