Normalizing speech transcriptions for Natural Language Processing

Morphosyntactic tagging and syntactic parsing are key parts of Natural Language processing. Many systems now reach exploitable results for written French texts (Véronis, 2000; Clément, 2001), but there were rare attempts to automatically annotate spoken textual data (see though Mertens, 2002; Valli et Véronis, 1999). Indeed, existing software are inadequate to analyse texts transcribed from speech and face specific problems, all related to the nature of the data: • for theoretical reasons (Blanche-Benveniste and Jeanjean, 1987), transcriptions of speech do not contain punctuation marks; nevertheless, most of the tools in Natural Language Processing are based on these marks in order to perform an initial segmentation of the text; • texts include meta information that does not need linguistic analysis (e.g. names of speakers, information on enunciation context