Natural language teaching takes significantly longer per input and yields significantly fewer teaching attempts and fewer features taught than sentence-selection¶

Claim · All claims
Evidence · 1 study (2 entries) · causal r1–r2 · q3
Currency · newest study 2025 · not yet checked against its sources · not reviewed by a person
Design limits · Rachel Love 2025: the intervention came bundled with another change · Rachel Love 2025 (2): the intervention came bundled with another change (read from the study by two models, 2026-10-27)

Subclaims¶

q3 i? Participants took significantly longer per teaching input (M = 73.0 s vs 28.8 s) and made significantly fewer teaching attempts (M = 12.4 vs 19.2) when teaching via natural language. → Rachel Love 2025 q3 i? Participants taught significantly fewer features via natural language (M = 11.0) than via sentence-selection (M = 13.6). → Rachel Love 2025 (2)

Evidence¶

Rachel Love 2025¶

Rachel Love, Edith Law, Philip R. Cohen, Dana Kulić. (2025). Teaching a Conversational Agent using Natural Language: Effect on Learning and Engagement. Int J Artif Intell Educ. https://doi.org/10.1007/s40593-025-00461-1

q3 · i? · causal · r2

Repeated measures ANOVA on teaching metrics in the 46-participant within-subjects study. "Participants took significantly longer to formulate their response when teaching using natural language" and made significantly fewer attempts, consistent with the need to read and type paraphrased responses.

"Participants took significantly longer to formulate their response when teaching using natural language ( F(1, 45) = 63.053, p <. 001), and made significantly fewer teaching attempts in total ( F(1, 45) = 105.361, p <. 001; natural lan- guage M = 12.4 (SD = 5.62); sentence-selection M = 19.2 (SD = 6.26))."

Rachel Love 2025 (2)¶

Rachel Love, Edith Law, Philip R. Cohen, Dana Kulić. (2025). Teaching a Conversational Agent using Natural Language: Effect on Learning and Engagement. Int J Artif Intell Educ. https://doi.org/10.1007/s40593-025-00461-1

q3 · i? · causal · r1

Same repeated measures ANOVA on teaching metrics. Fewer features were taught in the natural language condition, attributed by the authors to the fixed interaction length restricting the number of attempts possible.

"Participants also taught significantly fewer features when using natural language (F(1, 45) = 14.814, p <. 001; natural language M = 11.0 (SD = 4.10); sentence- selection M = 13.6 (SD = 4.40))."

Discussion¶

Related Claims¶

• A higher proportion of taught features is learned when teaching via natural language compared to sentence-selection — related
• Teaching via paraphrased natural language is perceived as more helpful for learning than sentence-selection, but this perception is not reflected in actual learning gains — related
• A higher proportion of untaught features is learned when teaching via sentence-selection compared to natural language — related
• The amount of paraphrasing users engage in during natural language teaching is positively related to improved learning outcomes — related