Sept 20, 2007:

Adaptive Training of Chinese Tones

Prof Chilin Shih, University of Illinois at Urbana-Champaign

In this talk, I will report the result of an experiment testing four treatments that train L2 learners to identify Mandarin lexical tones: (1) the control group with no tone training, (2) training with similar stimuli, (3) training with varied stimuli, and (4) training with varied stimuli through an adaptive training program, which is a computer program that chooses stimuli to match a student's level.

The adaptive group had the best performance where students showed an 8.2% improvement with less than three hours of training, where previous research reported using 40 hours of training.

We learned two lessons from the experiment: First, students responded differently to varied stimuli. While most students were helped by the varied input, some were confused and performed worse in the post-test. Adaptive learning alleviated the confusion.

Second, it is not straightforward to define "difficulty". We implemented an approximated, simplified, one-dimensional scale using speaker/listener distance as a guide, hypothesizing that speech projected to a distance is clearer, therefore easier to identify. This solution is at best an approximation. More importantly, difficulty is a multi-dimensional concept. Machine learning methods should be incorporated into adaptive programs to capture the complexity of the problem.