Section: Software and Platforms

The MICA parser

Participants : Benoît Sagot [correspondant] , Pierre Boullier.

Web page:

http://mica.lif.univ-mrs.fr/

MICA (Marseille-Inria-Columbia- AT&T) is a freely available dependency parser [48] currently trained on English and Arabic data, developed in collaboration with Owen Rambow and Daniel Bauer (Columbia University) and Srinivas Bangalore (AT&T). MICA has several key characteristics that make it appealing to researchers in NLP who need an off-the-shelf parser, based on Probabilistic Tree Insertion Grammars and on the Syntax system. MICA is fast (450 words per second plus 6 seconds initialization on a standard high-end machine) and has close to state-of-the-art performance (87.6% unlabeled dependency accuracy on the Penn Treebank).

MICA consists of two processes: the supertagger, which associates tags representing rich syntactic information with the input word sequence, and the actual parser, based on the Inria SYNTAX system, which derives the syntactic structure from the n-best chosen supertags. Only the supertagger uses lexical information, the parser only sees the supertag hypotheses.

MICA returns $n$-best parses for arbitrary $n$; parse trees are associated with probabilities. A packed forest can also be returned.