TY - GEN
T1 - Reliable estimates of interpretable cue effects with Active Learning in psycholinguistic research
AU - Einfeldt, Marieke
AU - Sevastjanova, Rita
AU - Zahner-Ritter, Katharina
AU - Kazak, Ekaterina
AU - Braun, Bettina
PY - 2021/3/8
Y1 - 2021/3/8
N2 - Studying the relative weighting of different cues for the interpretation of a linguistic phenomenon is a core element in psycholinguistic research. This research needs to strike a balance between two things: generalisability to diverse lexical settings, which requires a high number of different lexicalisations and the investigation of a large number of different cues, which requires a high number of different test conditions. Optimizing both is impossible with classical psycholinguistic designs as this would leave the participants with too many experimental trials. Previously we showed that Active Learning (AL) systems allow to test numerous conditions (eight) and items (32) within the same experiment. As stimulus selection was informed by the system’s learning mechanism, AL sped-up the labelling process. In the present study, we extend the use case to an experiment with 16 conditions, manipulated through four binary factors (the experimental setting and three prosodic cues; two levels each). Our findings show that the AL system correctly predicted the intended result pattern after twelve trials only. Hence, AL further confirmed previous findings and proved to be an efficient tool, which offers a promising solution to complex study designs in psycholinguistic research.
AB - Studying the relative weighting of different cues for the interpretation of a linguistic phenomenon is a core element in psycholinguistic research. This research needs to strike a balance between two things: generalisability to diverse lexical settings, which requires a high number of different lexicalisations and the investigation of a large number of different cues, which requires a high number of different test conditions. Optimizing both is impossible with classical psycholinguistic designs as this would leave the participants with too many experimental trials. Previously we showed that Active Learning (AL) systems allow to test numerous conditions (eight) and items (32) within the same experiment. As stimulus selection was informed by the system’s learning mechanism, AL sped-up the labelling process. In the present study, we extend the use case to an experiment with 16 conditions, manipulated through four binary factors (the experimental setting and three prosodic cues; two levels each). Our findings show that the AL system correctly predicted the intended result pattern after twelve trials only. Hence, AL further confirmed previous findings and proved to be an efficient tool, which offers a promising solution to complex study designs in psycholinguistic research.
U2 - 10.21437/Interspeech.2021-1524
DO - 10.21437/Interspeech.2021-1524
M3 - Conference contribution
BT - Proc. Interspeech 2021
ER -