Effects of Meaning-Preserving Corrections on Language Learning

Proceedings of the Fifteenth Conference on Computational Natural Language Learning, pages 97–105, Portland, Oregon, USA, 23–24 June 2011. c⃝2011 Association for Computational Linguistics Effects of Meaning-Preserving Corrections on Language Learning Dana Angluin ∗ Department of Computer Science Yale University, USA dana.angluin@yale.edu Leonor Becerra-Bonache Laboratoire Hubert Curien Universit´e de Saint-Etienne, France leonor.becerra@univ-st-etienne.fr Abstract We present a computational model of lan- guage learning via a sequence of interactions between a teacher and a learner. Experiments learning limited sublanguages of 10 natural languages show that the learner achieves a high level of performance after a reasonable number of interactions, the teacher can pro- duce meaning-preserving corrections of the learner’s utterances, and the learner can de- tect them. The learner does not treat correc- tions specially; nonetheless in several cases, significantly fewer interactions are needed by a learner interacting with a correcting teacher than with a non-correcting teacher. 1 Introduction A child learning his or her native language typically does so while interacting with other people who are using the language to communicate in shared situ- ations. The correspondence between situations and utterances seems likely to be a very important source of information for the language learner. Once a child begins to produce his or her own utterances, other people’s responses to them (or lack thereof) are another source of information about the lan- guage. When the child’s utterances fall short of adult-level competence, sometimes the other person in the conversation will repeat the child’s utterance in a more correct form. A number of studies have focused on the phenomenon of such corrections and questions of their frequency in child-directed speech ∗Research supported by the National Science Foundation, Grant CCF-0916389. and whether children can and do make use of them; some of these studies are discussed in the next sec- tion. In this paper we construct a computational model with a learner and a teacher who interact in a se- quence of shared situations. In each situation the teacher and learner interact as follows. First the learner uses what it has learned about the language to (attempt to) generate an utterance appropriate to the situation. The teacher then analyzes the correct- ness of the learner’s utterance and either generates an utterance intended as a correction of the learner’s utterance, or generates another utterance of its own appropriate to the situation. Finally, the learner uses information given by its own utterance, the teacher’s utterance and the situation to update its knowledge of the language. At the conclusion of this interac- tion, a new interaction is begun with the next situa- tion in the sequence. Both the learner and the teacher engage in com- prehension and production of utterances which are intended to be appropriate to their shared situation. This setting allows us to study several questions: whether the teacher can offer meaningful correc- tions to the learner, whether the learner can detect intended corrections by the teacher, and whether the presence of corrections by the teacher has an ef- fect on language acquisition by the …

Content truncated to 500 words