Graphics Reference
In-Depth Information
of spoken French. In a second stage, the output of the parser was
manually corrected for the totality of the CID. The annotation process
is time-consuming whether it is manual or automatic. The manual
annotation requires several annotators (either expert or not, sometimes
both) and tests of labeling consistency to measure inter-annotator
agreement. The automatic annotation is less time-consuming but also
requires evaluation between the different tools or involves manual
corrections, which enable to evaluate the performance of the parser
(only 5% of error rates).
2.4 Prosodic annotations
The prosodic level can be annotated in a manual or an automatic
way depending on whether we observe rather phonological (more
abstract) phenomena or phonetic parameters. In OTIM, we focused
on the prosodic phrasing which corresponds to the structuring of
speech material in terms of boundaries and groupings. The manual
annotation was very time-consuming but was necessary to improve
the knowledge of prosodic domains in French. In a first stage, such a
manual annotation made by experts enabled us to test the robustness
of annotation criteria. A previous study involved two experts; results
have shown a very good inter-coder agreement and kappa score for the
higher level of constituency (IP) (Nesterenko et al., 2010). In a second
stage, the elaboration of a guideline for transcribing prosodic units
in French by naïve annotators enabled us to test the reduplicability
of these annotation criteria. Naïve transcribers have to annotate four
levels of prosodic break defined in terms of a ToBI-style annotation 3 (0
= no break; 1 = AP break; 2 = ip break; 3 = IP break) in Praat (Boersma
and Weeninck, 2009). The global aim is to develop a phonologically
based transcription system for French that would be consistent enough
to be amenable to automatic labeling. One of the steps is to compare
the manual annotations. Another step consists in improving existing
automatic tools (such as Intsint for example, Hirst et al., 2000) by
comparing the output of different annotation tools and manual expert/
naïve annotation (Peskhov et al., 2012).
At last, another aspect of prosodic annotation concerns the
intonation contours associated to intonational or intermediate phrases
(levels 3 and 2 above). Pitch contours are formally and functionally
defined (Portes et al. (2007) for details). Intonation contours were
coded for six speakers.
3 http://www.ling.ohio-state.edu/~tobi/ame_tobi/annotation_conventions.html
 
Search WWH ::




Custom Search