Retrodiction as Delayed Recurrence: the Case of Adjectives in Italian and English

Retrodiction as Delayed Recurrence: the Case of Adjectives in Italian and English Raquel G. Alhama1,2 Francesca Zermiani 2,3 Atiqah Khaliq2 1Tilburg University, Department of Cognitive Science & Artificial Intelligence, Tilburg, The Netherlands 2Max Planck Institute for Psycholinguistics, Language Development Department, Nijmegen, The Netherlands 3University of Stuttgart, Department of Teaching and Learning with Intelligent Systems, Stuttgart, Germany rgalhama@tilburguniversity.edu, francesca.zermiani@ife.uni-stuttgart.de, atiqah.khaliq@mpi.nl Abstract We address the question of how to account for statistical dependencies in an online pro- cessing account of human language acquisi- tion. We focus on descriptive adjectives in English and Italian, and show that the acqui- sition of adjectives in these languages likely relies on tracking both forward and backward regularities. Our simulations confirm that forward-predicting models like standard Re- current Neural Networks cannot account for this phenomenon due to the lack of backward prediction, but the addition of a small delay (as proposed in Turek et al., 2019) endows the RNN with the ability to not only predict but also retrodict. 1 Introduction Sensitivity to statistical regularities allows for effi- cient lexical processing. As a sentence unfolds, the experienced words convey information that humans use to anticipate upcoming words, and gain thereby processing speed. This has been evidenced in a long tradition of studies with human reading data, which reveal that words that are more predictable given their context are more likely to be read faster or even skipped (Ehrlich and Rayner, 1981). The ability to track statistical regularities during language processing is present at a very young age, and can be recruited for language learning. Be- fore their first birthday, infants are able to use this skill to identify words in unknown languages cre- ated with artificial (Saffran et al., 1996; Aslin et al., 1998) or natural words (Pelucchi et al., 2009b), demonstrating that this ability is useful for learning language-like stimuli (see Saffran, 2020 for a re- view). Studies have found that, before their second year of age, toddlers already engage in predictive processing to identify familiar words before they are complete (Swingley et al., 1999; Fernald et al., 2001), and are capable of anticipating upcoming words (Fernald and Hurtado, 2006; Lew-Williams and Fernald, 2007). Given this relation between online processing and learning, it is perhaps unsurprising that chil- dren with more efficient lexical processing are also those with faster vocabulary growth (Fernald et al., 2006; Fernald and Marchman, 2012; Weisleder and Fernald, 2013; Donnelly and Kidd, 2020). From a cross-linguistic perspective, this suggests that ty- pological variation on the statistical regularities of different languages should be either equally tracked during processing, or reflected in cross-linguistic differences in learning. In our work, we focus on one such typological feature: in particular, word order of descriptive ad- jectives in English (which occur pre-nominally), and Italian (which appear mostly post-nominally, but also pre-nominally). We first show that this difference in word order bears a different pattern of statistical dependencies in these languages, related to the direction in which the words in these con- structions are more predictable (forward in …

Content truncated to 500 words