A MDL-based Model of Gender Knowledge Acquisition

CoNLL 2008: Proceedings of the 12th Conference on Computational Natural Language Learning, pages 73–80 Manchester, August 2008 A MDL-based Model of Gender Knowledge Acquisition Harmony Marchal1, Benoît Lemaire2, Maryse Bianco1, and Philippe Dessus1 1L.S.E. and 2Laboratoire TIMC-IMAG University of Grenoble, FRANCE <first name>.<last name>@upmf-grenoble.fr Abstract This paper presents an iterative model of knowledge acquisition of gender infor- mation associated with word endings in French. Gender knowledge is represented as a set of rules containing exceptions. Our model takes noun-gender pairs as in- put and constantly maintains a list of rules and exceptions which is both coher- ent with the input data and minimal with respect to a minimum description length criterion. This model was compared to human data at various ages and showed a good fit. We also compared the kind of rules discovered by the model with rules usually extracted by linguists and found interesting discrepancies. 1 Introduction In several languages, nouns have a gender. In French, nouns are either masculine or feminine. For example, you should say le camion (the truck) but la voiture (the car). Gender assignment in French can be performed using two kinds of information. Firstly, lexical information, related to the co-occurring words (e.g., articles, adjec- tives) which most of times marks gender unam- biguously. Secondly, sublexical information, es- pecially noun-endings, are pretty good predictors of their grammatical gender (e.g., almost all nouns endings in –age are masculine). Several word endings can be used to reliably predict gender of new words but this kind of rules is never explicitly taught to children: they have to implicitly learn that knowledge from exposure to noun-gender pairs. It turns out that children as young as 3 already constructed some of these © 2008. Licensed under the Creative Commons Attri- bution-Noncommercial-Share Alike 3.0 Unported license (http://creativecommons.org/licenses/by-nc- sa/3.0/). Some rights reserved. rules, which can be observed by testing them on pseudo-words (Karmiloff-Smith, 1979). This paper presents an iterative model of the way children may acquire this gender knowl- edge. Its input is a large random sequence of noun-gender pairs following the distribution of word frequency at a given age. It is supposed to represent the words children are exposed to. The model constantly maintains a list of rules and exceptions both coherent with the input data and minimal with respect to an information theory criterion. This model was compared to human data at various ages and showed a good fit. We also compared the kind of rules discovered by the model with rules usually extracted by lin- guists and found interesting discrepancies. 2 Principle of Simplicity Gender knowledge is learned from examples. Children are exposed to thousands of nouns which are most of the time accompanied with a gender clue because of their corresponding de- terminer or adjective. For instance, when hearing “ta poussette est derrière le fauteuil” [your stroller is behind the armchair], a child knows that poussette is feminine because of the femi- nine possessive determiner ta, and that fauteuil is masculine because of the masculine determiner le. …

Content truncated to 500 words