The Integration of a Part - of - Speech Taggerinto

We describe how part-of-speech information delivered by a tagger (the mpro tool) has been integrated into the alep (Advanced Language Engineering Platform) system. For this we extended an approach described within the ls-gram project, which consisted in deening the Text Handling component of alep in such a way that so-called \messy details" are handled within this subsystem, hence keeping the (linguistic) parser free from such tasks. We just extended the tagging strategy used for this purpose to normal words and modiied the default tagging of words proposed by the alep system in order to incorporate informationdelivered by the part-of-speechtagger. The resulting tagging is converted by means of \lift" rules into partial linguistic descriptions, which provide the direct input to the grammatical analysis. We show that this procedure substantially reduces the parse times of the system. 1 Background The starting point of our work has been the treatment of so-called`messy details' within the alep platform 2. Messy details are text constructs which do not lend themselves well to treatment by traditional techniques for linguistic analysis, whence their`messiness'. Typical examples are numbers, codes or other (sequences of) word-forms which can occur in many variations, making impossible a comprehensive treatment by traditional means. Word level phenomena are usually the most frequently occurring messy details, including such things as dates, numbers and proper names. For any realistic NLP application these types of constructs must be processed eeciently, the alternative being coding them individually in some lexicon and/or implementing sets of grammar rules for parsing them syntactically. This problem area was given priority in the ls-gram project 3 which aimed to integrate an approach to messy details into a large-scale grammar implementation. How we dealt with this phenomena has been described in 2 short papers 4. The alep platform provides the user with a Text Handling (TH) component which allows a \prepro-cessing" of inputs. Input texts will rst go through a processing chain consisting in a sgml-based tagging of their elements. The default setup of the system deenes the following processing chain: the text is rst converted to an EDIF (Eurotra Document Interchange Format) format. Then a subsequent recognition process is provided: paragraph recognition, sentence recognition and word recognition. The 1 We would like to thank the anonymous reviewers for their precious comments on our contributions for this workshop.