Learning Chinese Syllables by Attending to

Auditory, Visual, and Lexical Cues for Tone

Jasmine Kwasa (CMU Lab in Multisensory Neuroscience)

Mentor: Philip Pavlik (U Memphis)

Background: Chinese Tones

Mandarin Chinese has 5 tones that distinguish meaning between syllables of the same spelling.

​

These tones can be represented as:

• A tone number (1-5)
• Pinyin, a Roman spelling system
• Pitch contours, an approximation of the auditory frequency

​

Learners who speak non-tonal languages (like English) tend to struggle with categorizing these tones in an auditory context as compared to learners who speak tonal languages (like Thai).

​

Background: Dataset Information

This 2011 study asked if presenting pinyin, contours, or the tone numbers in an online feedback tutor would help students learn these tones better.

Learning a Tonal Language by Attending to the Tone: An In Vivo Experiment

Ying Liu (UPitt), Min Wang, Charles Perfetti, Brian Brubaker, Sumei Wu, Brian MacWhinney (CMU)

​

​

Intelligent Tutor Setup

Methods: The investigators assigned each student to one of three training conditions providing different kinds of sensory information for the response. Students listen to a one- or two-syllable word and try to identify the tone(s).

​

​

Condition 1:

Visual + Lexical

Condition 2

Lexical Only

Condition 3

Visual Only

Our Questions and Hypotheses

​

​

• Are the error rates the same for each tone in each condition?
• We expect tones 2 and 3 to have slower learning rates for non-tonal language speakers, based on our prior knowledge
• Is there a difference in the lesson learning rates for each condition?
• 8 total “lessons” across semester, each with increasing number of syllables to learn
• Previous article indicates that Pinyin + Contour is fastest
• Is there a difference in the opportunity learning rates for each condition?
• We think this is a more sensitive measure of learning since opportunity, the number of exposures to a certain concept, is more granular than a lesson

​

Results: Differences in KC

Data is so noisy (!), especially for the neutral tone

Our model indicates that tones 2, 3, and neutral were at the hardest difficulty level

Raw data

Our AFM Model

Our model:

ln(p/1-p) = b0 + b1*studentID

+ b2*KC + b3*KC*Opportunity

KC = knowledge component (5 tones to be learned)

Opportunity = number of exposures the student has had to the KC

studentID = identification for each individual

Is there a difference in the lesson learning rates for each condition?

​

Results: Differences in Condition

No significant differences found between the article and our analysis

​

This might not be the best way to visualize learning rates since there were only 8 lessons across the semester, each with larger syllable sets to learn

​

From article

Our analysis

Is there a difference in the opportunity learning rates for each condition?

ln(p/1-p) = b0 + b1*studentID

+ b2*KC + b3*KC*Opportunity

ln(p/1-p) = b2*KC + b3*KC*Opportunity

KC = knowledge component (5 tones to be learned)

Opportunity = number of exposures the student has had to the KC

studentID = identification for each individual

NO differences found!

Comparing Models: Lessons vs Opportunities

From article

From our analysis

NO differences found!

Conclusions / Discussion

• Was one learning rate faster than the others?
• Based on confidence interval comparisons, NO
• Initial difficulty was definitely different (tones 1 and 4 were easiest)
• Do these results align with the authors’ results?
• No, but they are similar
• Analyzing learning data sets using Opportunity counts instead of higher-level grouping like Lessons is preferred.

​

Results from article. Perhaps their conclusion (that the Contour + Pinyin condition had the highest learning rate) was due to ceiling effects and other confounding factors

Limitations and Recommendations

The conditions might not have been separated according to clearly defined multisensory processes.

Analysis for each syllable-level KC would have been more revealing of the challenges Chinese-learners face

Recommendations:

• Give this tutor to completely naïve learners (e.g., Amazon Turk)
• The intercepts (difficulty) indicate that there was previous knowledge of the tones, perhaps from class instruction
• Account for dropout
• Account for each individuals’ learning rate (using Performance Factors Analysis, for example)

References, Thank You’s, etc.

Tone Graphic: https://www.iwillteachyoualanguage.com/learn/chinese/chinese-tips/demystifying-chinese-tones

Ying Liu, Min Wang, Charles Perfetti, Brian Brubaker, Sumei Wu, Brian MacWhinney, “Learning a Tonal Language by Attending to the Tone: An In VIvo Experiment” 2011 Language Learning

​

Thank you to John Stamper, Ken Koedinger, and Phil Pavlik for your mentorship on this project!

​

​

​

​

​