Disfluency

In oral language, speakers form sentences while speaking, which means that sometimes they might search for words while speaking. They may repeat a word until they find the right one, or they may try with one word and then decide to change it. In SUD, we use two relations for the annotations of this contructions:

• reparandum for repetition
• conj:reform for reformulation

Note

Up to version 2.18, the relation conj:dicto was used as a proxy for both repetitions and reformulations. Some SUD treebanks were not updated and still use conj:dicto even if the relation is deprecated for new annotations.

The next example shows French sentence with both annotations.

The conj:reform relation applies also when the linked words are not the same.

Unfinished constructions

Sometimes speakers utter a half-finished construction. In this case, a word cannot be attached to its head, because the speaker decided not to say it.

Below we can see an example when a speaker starts with one unfinished word m~, then decides it doesn’t fit and searches for a more fitting word virus.

Warning

In previous versions, the deep features @scrap was used for unfinished constructions. This is now deprecated and should be used in SUD annotations.