Are statistics-based approaches good enough for NLP? A case study of maximal-length NP extraction in Mandarin Chinese

Statistics-based approaches became very popular in recent NLP researches, because of their apparent advantages over linguistics or rule-based approaches. Some even claimed that it would not be necessary to employ the latter approach at all. Thus, it seemed necessary to evaluate such claim and the applicability of the former to NLP in general. Because of the usefulness of noun phrases (NPs) in many applications, in this paper, we present a simple statistics-based partial parser to detect the boundaries of maximal-length NPs in part-of-speech tagged Chinese texts. On the basis of our experimental results, we will show that statistics-based approaches with purely part-of-speech tags are not adequate for NP extraction in Chinese; they fail to handle cases with structural ambiguity. Our experiments suggest that syntactic and semantic checking is necessary to correctly mark the boundary of maximal-length NPs in Chinese. We conclude with possible solutions to the problematic cases for statistics-based approaches.