Backchannels are short productions uttered by one participant while another speaker holds the floor.
To be recognized as backchannels, such utterances (i) must be addressed to the content of the other speaker’s contribution; (ii) must not be required or expected by the preceding turn (e.g. answers to wh- and polar questions are expected and required, so they cannot be considered backchannels); (iii) must not require a reaction from the main speaker.
The backchannel feature is encoded as such: Backchannel:⟨sent_id⟩::⟨tok_id⟩, on the root of the sentence where the backchannel occurs. The sent_id is that of one of the previous sentences. Most often it is the previous sentence, but because of how speech is segmented, it can be the sent_id of a sentence further back. The tok_id is generally that of the previous sentence’s root. In some edge cases, when the previous sentence is rather long and there are multiple utterances that can be annotated as backchannels, we differentiate the tok_id of each annotation, using the tok_id of the head of the last completed segment.
Base case
Edge case 1
Here is a long sentence with two backchannels, each with a different tok_id.
Dependency based
This annotation is used in the dependency based view, merging backchannels with their corresponding head sentences.
Note
This kind of annotation is only available in specific dependency based corpora like SUD_FRANCH-Rhapsodie@db.
Example 3: Dependency based view