Fourtassi, A. 1, 2 & Dupoux, E. 2, 3
1 Laboratoire de Sciences Cognitives et Psycholinguistique
2 Ecole Normale Supérieure
3 Ecole des Hautes Etudes en Sciences Sociales
How do children learn the words of their native language amidst high ambiguity? A well studied strategy (cross situational learning) narrows down the possible referents of a word by keeping track of its exposure statistics across situations. We propose an additional strategy where the learner may guess the referent of a word only using its co-occurrence with known words, and no prior exposure to any informative situation.
The feasibility of such ?zero-shot learning? has been demonstrated in machine learning, but has not been tested experimentally in humans. In a web-based study, participants were first trained to learn the pairing of 4 words in an artificial language with 4 objects. The objects belonged to the category of animals or vehicles. Then, they listened to 'sentences' made of words from this artificial language. Each sentence contained the words from one category plus a new word that consistently co-occur with them. In the test phase, participants were asked to infer the meaning of the new words, by choosing between a new animal or a new vehicle. The results confirmed our prediction: participants were significantly above chance level in predicting the meaning of test words only based on their patterns of co-occurrence.