Incorporating Head Recognition into a CRF Chunker

摘要

While rule-based shallow parsers usually recognise phrases' syntactic heads, the same does not hold for statistical syntactic chunkers. The task of finding heads within already recognised chunks is not trivial for freer word order languages like German or Polish, while this information may be very useful. We propose a simple solution that allows to incorporate head recognition into existing chunkers by extending the standard IOB2 representation with information on head location. To evaluate this approach we introduced the new representation into a CRF chunker for Polish. Although this idea is very simple, the results are surprisingly good.