A "not-so-shallow" parser for collocational analysis

Abs t rac t . Collocational analysis is the basis of many s tud ies on lexical acquis i t ion . Collocations are extracted from corpora using more or less shallow processing techniques, that span from purely statistical methods to partial parsers. Our point is that, despite one of tile objectives of collocational analysis is to acquire high-coverage lexical data at low human cost, this is often not the case. Human work is in fact requ i red for the init ial t ra in ing of most s tat is t ical ly based methods. A more serious problem is that shallow processing techniques produce a noise that is not acceptable for a fully automated system. We propose in this paper a not-so-shallow parsing strategy that reliably detects binary and ternary relations among words. We show that adding more syntactic knowledge to the. recipe significantly improves the recall and precision of tile detected collocations, regardless of any subsequent statistical computation, while still nleet ing the cornputat ional requi, 'ements of corpus parsers.