Morpho-syntactic labelling of an oral corpus by decomposing labels

A morpho-syntactic tagger associates to each word of a corpus a label which recapitulates its morpho-syntatic properties in the text. In corpora from oral data, not only do we have to face the usual problem of multi-labels words, but also the more specific problems of disfluences (repetitions, ungrammatical constructions...), of non existing words and of the lack of punctuation marks [1]. First, inspired by the Cordial tagger, we have defined a new set of morphosyntactic labels well adapted to oral corpora. These labels are hierarchically defined according to three different levels: a POS level (L0, with 16 different labels), a morphological level (L1, with 72 labels) and a syntactico-semantic level (L3, with 107 labels). Then, we have built a reference corpus for our new set of labels. It has been obtained by using Cordial, then modifying its labels by scripts and manual corrections. The learning corpus finally contains 18 500 words belonging to 1 750 distinct “sequences”.