Multilingual Models for the Rainbow Parser


The Rainbow Parser (or RParser) is a phrase-structure syntactic parser developed at the University of Edinburgh by the informal research group Cohort. At its core, the use of a latent-variable PCFG model. Its training procedure is based on spectral methods of learning. The parser is not publicly available yet. However, if you are interested in using it for your research, contact Shay Cohen (scohen AT inf.ed.ac.uk) or Shashi Narayan (snaraya2 AT inf.ed.ac.uk).

Click for the following paper.


@inproceedings{narayan-16b,
  title={Optimizing Spectral Learning for Parsing},
  author={Shashi Narayan and Shay B. Cohen},
  booktitle={Proceedings of {ACL}},
  year={2016}
}


Below we include the table of results on the test sets from the SPMRL shared task to parse morphologically rich languages. For a legend, see the paper (Tables 2 and 3).


LanguageCL van.CL opt.SP van.SP opt.Berkeley
  Basque 79.6 81.4 79.9 80.5 74.7
  French 74.3 75.6 78.7 79.1 80.4
  German (NEGRA) 76.4 78.0 78.4 79.4 80.1
  German (TiGeR) 74.1 76.0 78.0 78.2 78.3
  Hebrew 86.3 87.2 87.8 89.0 87.0
  Hungarian 86.5 88.4 89.1 89.2 85.2
  Korean 76.5 78.4 80.3 80.0 78.6
  Polish 90.5 91.2 91.8 91.8 86.8
  Swedish 76.4 79.4 78.4 80.9 80.6

Results are given using EVALB with the parameters file from the SPMRL shared task. Click the numbers in blue to download the respective L-PCFG model.