Computerized training with AI and human voices facilitates L2 prediction of grammatical gender agreement
Document Type
Paper Presentation
Presenter Language
English
Research Area
Psycholinguistics, Second language acquisition, Syntax
Location
MBSC Dodge Room 302
Start Date
19-10-2024 10:00 AM
End Date
19-10-2024 10:30 AM
Abstract
Artificial Intelligence (AI) voice generators have gained popularity due to the rapid development of AI technology. However, little is known about the differences in L2 learning between human and artificial voices. This study investigates the effects of computerized training using human or artificial voices on the prediction of grammatical gender agreement by beginning L2 learners of a genderless L1. Monolinguals and advanced L2 learners use lexical and syntactic cues to generate correct predictions during gender agreement computation, but beginning L2 learners underuse lexical cues (Grüter et al., 2012; Hopp, 2016) and struggle making correct predictions. Computerized training facilitated the prediction of verb suffixes based on tones (Schremm et al., 2017) and may also improve the prediction of noun suffixes based on gender lexical cues (a noun’s inherent gender) and syntactic cues (gender-marked articles and suffixes).
Seventy-seven beginning English learners of Spanish played a digital game with an AI (n = 39) or a human voice (n = 38) for ten minutes per day for eight days. Participants listened to an article-noun pair (el jamón ‘the ham’), moved a frog to a masculine (sabroso ‘deliciousM’) or a feminine (sabrosa ‘deliciousF’) platform, and received feedback (correct/incorrect + correct utterance). Each participant heard 384 determiner-noun pairs, 96 per condition: [+/- gender-marked article, +/- gender-marked noun]: el queso ‘theM cheeseM’; su queso ‘his/hercheeseM’; el jamón ‘theM hamM’; su jamón ‘his/herhamM’.
GLMMs revealed no differences between AI and human voices in accuracy or RTs. This result suggests that AI and natural speech are equally beneficial for L2 grammar learning, and that AI can help create time-saving and cost-effective materials for L2 learning and language experiments. Also, training decreased RTs in all conditions and increased accuracy in all conditions except the condition without gender-marked articles or nouns (su jamón). No accuracy gains in the condition lacking syntactic gender cues indicate that computerized training is effective to learn syntactic, but not lexical, gender cues. This finding supports L2 models claiming that L2 learners’ difficulty acquiring grammatical gender agreement results from a weak and unstable knowledge of lexical gender (Grüter et al., 2012; Hopp, 2013, 2016).
Keywords: AI, digital game, computerized training, gender agreement, online learning
Computerized training with AI and human voices facilitates L2 prediction of grammatical gender agreement
MBSC Dodge Room 302
Artificial Intelligence (AI) voice generators have gained popularity due to the rapid development of AI technology. However, little is known about the differences in L2 learning between human and artificial voices. This study investigates the effects of computerized training using human or artificial voices on the prediction of grammatical gender agreement by beginning L2 learners of a genderless L1. Monolinguals and advanced L2 learners use lexical and syntactic cues to generate correct predictions during gender agreement computation, but beginning L2 learners underuse lexical cues (Grüter et al., 2012; Hopp, 2016) and struggle making correct predictions. Computerized training facilitated the prediction of verb suffixes based on tones (Schremm et al., 2017) and may also improve the prediction of noun suffixes based on gender lexical cues (a noun’s inherent gender) and syntactic cues (gender-marked articles and suffixes).
Seventy-seven beginning English learners of Spanish played a digital game with an AI (n = 39) or a human voice (n = 38) for ten minutes per day for eight days. Participants listened to an article-noun pair (el jamón ‘the ham’), moved a frog to a masculine (sabroso ‘deliciousM’) or a feminine (sabrosa ‘deliciousF’) platform, and received feedback (correct/incorrect + correct utterance). Each participant heard 384 determiner-noun pairs, 96 per condition: [+/- gender-marked article, +/- gender-marked noun]: el queso ‘theM cheeseM’; su queso ‘his/hercheeseM’; el jamón ‘theM hamM’; su jamón ‘his/herhamM’.
GLMMs revealed no differences between AI and human voices in accuracy or RTs. This result suggests that AI and natural speech are equally beneficial for L2 grammar learning, and that AI can help create time-saving and cost-effective materials for L2 learning and language experiments. Also, training decreased RTs in all conditions and increased accuracy in all conditions except the condition without gender-marked articles or nouns (su jamón). No accuracy gains in the condition lacking syntactic gender cues indicate that computerized training is effective to learn syntactic, but not lexical, gender cues. This finding supports L2 models claiming that L2 learners’ difficulty acquiring grammatical gender agreement results from a weak and unstable knowledge of lexical gender (Grüter et al., 2012; Hopp, 2013, 2016).
Keywords: AI, digital game, computerized training, gender agreement, online learning