Math Learning From a Gesturing Avatar

• In Event: 33.078 - Roundtable Session 11
  In Roundtable Session: 33.078-16 - The Learning at Hand: Gesture Production in Virtual Pedagogical Agents

Abstract

Purpose
One important question is whether gestures produced by embodied pedagogical agents will facilitate math learning comparable to spontaneously produced instructor gesture (Cook, 2011; Goldin-Meadow & Alibali, 2013).

Perspective
We present an experimental study manipulating gesture using an instructor avatar implemented as a computer animation character. With animation, we can create complex stimuli accurately and efficiently while controlling nonverbal and verbal parameters that are not of interest.

Methods
Fifty-eight 9-10 year old children participated in an individual, computerized tutorial with an avatar (Figure 5) who either gestured or did not gesture.

Children first completed a pretest of six equal addends mathematical equivalence problems. Then children viewed an introductory statement about equivalence. They then viewed six explanations of novel, equal addends equivalence problems. Explanations were based on those previously used to show beneficial effects of gesture during instruction, both live (e.g. Singer & Goldin-Meadow, 2005), and on video (e.g. Cook, Duffy & Fenn, 2013). Matched pairs were created, with one explanation in each pair containing hand gesture. After each explanation, children solved an equal addends equivalence problem. Finally, children completed a posttest, a transfer test consisting of four equivalence problems without equal addends, and a generalization test consisting of six true/false questions about the concept of equality (adapted from Matthews et al., 2012).

Data
We analyzed performance for children who did not solve any pretest problems correctly (N=28). A multilevel logistic regression model accounted for variability across individual subjects as well as variability in problem difficulty. The log of the odds of correctly solving each problem was predicted from condition interacting with test. The maximal random effects structure was included by subject along with a random problem intercept.

Results
There was an effect of test, with the posttest and transfer tests more difficult than the true/false generalization test (Posttest=-2.46, z=1.94, p=.053; Transfer=-2.58, z=1.54, p=.12). There was a significant effect of condition, with children in the gesture condition performing better than children in the no gesture condition (Gesture=3.39, z=2.80, p<.01, see Figure 6). There were no significant interactions between condition and test (all p’s>.37). On average, children in the gesture condition were correct on 88% of problems after training, compared to 72% for the no gesture condition.

We also examined performance on the transfer and generalization tests for only children who were successful on the posttest (N=22). There was again an effect of gesture condition (Gesture =2.83, z=2.23, p=.026), in favor of the gesture condition, and no interaction (p>.16).

Significance
These findings confirm that children can benefit from gesture produced by animated pedagogical agents. Moreover, these findings extend prior work on the role of gesture in math learning to reveal that instruction that includes gesture can influence children’s conceptual knowledge. Finally, these findings provide evidence that the beneficial effect of gesture is not due to differences in eye gaze, prosody, or body position; these factors were rigorously controlled across conditions, which is a unique strength of our computer animation approach to stimuli creation.

Authors

• Susan Wagner Cook, University of Iowa

• Voicu Popescu, Purdue University