Title:	Human and Transformer-Based Prosodic Phrasing in Two Speech Genres
Authors:	Volín, Jan Řezáčková, Markéta Matoušek, Jindřich
Citation:	VOLÍN, J. ŘEZÁČKOVÁ, M. MATOUŠEK, J. Human and Transformer-Based Prosodic Phrasing in Two Speech Genres. In 23rd International Conference, SPECOM 2021, St. Petersburg, Russia, September 27–30, 2021, Proceedings. Cham: Springer, 2021. s. 761-772. ISBN: 978-3-030-87801-6 , ISSN: 0302-9743
Issue Date:	2021
Publisher:	Springer
Document type:	konferenční příspěvek ConferenceObject
URI:	2-s2.0-85116335966 http://hdl.handle.net/11025/47258
ISBN:	978-3-030-87801-6
ISSN:	0302-9743
Keywords in different language:	Phrasing;Prosodic boundaries;Prediction;Syntactic constituents;Transformers;T5
Abstract in different language:	The chief objective of the study was to observe phrasing behaviour of transformer-based neural networks from the linguistic point of view. The transformer-based architecture mapped prosodic phrasing in isolated sentences read out on request, but was commanded to predict prosodic phrases in continuous texts of journalistic style taken from radio news bulletins. The transfer was quite successful in that most of the prosodic phrase boundaries in the actual newsreading (established by expert auditory analysis) were correctly suggested by the machine. This result is not unexpected as both genres belong to clearly enunciated informative speaking style. The outcome partially rehabilitates the so-called laboratory speech, which is sometimes branded as ecologically invalid. The follow-up analyses revealed that the differences between human phrasing in news bulletins and the partition suggested by the machine can be classified into meaningful linguistic categories based on the syntactic structure or semantic contents, and as such, they can inform further research design.
Rights:	Plný text je přístupný v rámci univerzity přihlášeným uživatelům. © Springer
Appears in Collections:	Konferenční příspěvky / Conference Papers (KKY) OBD

Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.