Abstract
The relationship between gender and word ending in French is a quasiregular one (e.g., most words ending in -ette are feminine, but not all). As such, the gender of low-frequency irregular forms (e.g., squelette, which is masculine) should take longer to classify than low-frequency regular forms according to neural network models. A regularity effect was found in Experiment 1, but it did not interact with word frequency. It was further revealed that there was difficulty in making gender decisions (Experiment 2) and gender verification responses (Experiment 3) to words whose endings were highly informative of gender, but whose associated article was not. These words were place names beginning with a vowel, like Australie, which do not take an indefinite article and whose definite article is ambiguous (1'). How a neural network might handle these results is discussed, and an alternative account is considered whereby there are two potential sources of gender information, lexical and nonlexical, with the latter being used to confirm the former.