This article presents the first collection of French Winograd Schemas. Winograd Schemas form anaphora resolution problems that can only be resolved with extensive world knowledge. For this reason the Winograd Schema Challenge has been proposed as an alternative to the Turing Test. A very important feature of Winograd Schemas is that it should be impossible to resolve them with statistical information about word co-occurrences: they should be Google-proof. We propose a measure of Google-proofness based on Mutual Information, and demonstrate the method on our collection of French Winograd Schemas.

While many studies have shown that toddlers are able to detect syntactic regularities in speech, the learning mechanism allowing them to do this is still largely unclear. In this article, we use computational modeling to assess the plausibility of a context-based learning mechanism for the acquisition of nouns and verbs. We hypothesize that infants can assign basic semantic features, such as “is-an-object” and/or “is-an-action,” to the very first words they learn, then use these words, the semantic seed, to ground proto-categories of nouns and verbs. The contexts in which these words occur, would then be exploited to bootstrap the noun and verb categories: unknown words are attributed to the class that has been observed most frequently in the corresponding context. To test our hypothesis, we designed a series of computational experiments which used French corpora of child-directed speech and different sizes of semantic seed. We partitioned these corpora in training and test sets: the model extracted the two-word contexts of the seed from the training sets, then used them to predict the syntactic category of content words from the test sets. This very simple algorithm demonstrated to be highly efficient in a categorization task: even the smallest semantic seed (only 8 nouns and 1 verb known) yields a very high precision (~90% of new nouns; ~80% of new verbs). Recall, in contrast, was low for small seeds, and increased with the seed size. Interestingly, we observed that the contexts used most often by the model featured function words, which is in line with what we know about infants' language development. Crucially, for the learning method we evaluated here, all initialization hypotheses are plausible and fit the developmental literature (semantic seed and ability to analyse contexts). While this experiment cannot prove that this learning mechanism is indeed used by infants, it demonstrates the feasibility of a realistic learning hypothesis, by using an algorithm that relies on very little computational and memory resources. Altogether, this supports the idea that a probabilistic, context-based mechanism can be very efficient for the acquisition of syntactic categories in infants.

The Corpus for Idiolectal Research (CIDRE) is a collection of fiction works from 11 prolific 19th-century French authors (4 women, 7 men; 22-62 works/author; total of 37 million words). Every work is dated with the year it was written. Using programming scripts, the works have been gathered from open source platforms, for example La Bibliothèque électronique du Québec, and stripped of paratext (text not being part of the novel, e.g. prefaces). We distribute the text files, the dating, other metadata and the programming scripts under an open source license. CIDRE is the first resource of French for the study of style and idiolect in a diachronic manner (i.e. stylochronometry) on a larger scale.