The adaptable speaker: A theory of implicit learning in language production.

Abstract

The language production system continually learns. The system adapts to recent experiences while also reflecting the experience accumulated over the lifetime. This article presents a theory that explains how speakers implicitly learn novel phonotactic patterns as they produce syllables. The learning is revealed in their speech errors. For example, if speakers produce syllable strings in which the consonant /f/ is always a syllable onset, their slips will obey this rule; /f/'s will then slip mostly to onset positions. The article reviews over 30 phenomena related to this finding. To explain phonotactic learning, the article presents four linked "mini-theories," each of which addresses components of the data. The first mini-theory, the production theory, provides an account of how speech errors arise during the assembly of word forms. The second, the learning theory, characterizes the implicit learning of phoneme distributions within the production system. The third mini-theory, the consolidation theory, augments the learning theory to explain instances in which this learning depends on a period of time, possibly a sleep period, before it is expressed. The final mini-theory, the developmental theory, addresses cases in which learning varies between children and adults, and depends on speakers' early linguistic experience. The resulting theory forges links between these diverse aspects of psychology.

Surprisal does not explain syntactic disambiguation difficulty: evidence from a large-scale benchmark

“…A central factor invoked to explain this fact is prediction: by anticipating upcoming words, readers are enabled to rapidly integrate them into their interpretation of the sentence (Kutas et al, 2011). This explanation fits with the growing evidence for prediction as an organizing principle of linguistic cognition in particular (Dell et al, 2021;Pickering and Garrod, 2013) and the brain more generally (Bar, 2007). In parallel, much recent work has shown that language modelscomputational systems trained to predict the next word in a sentence-serve as a powerful foundation for language understanding by computers (Peters et al, 2018;Brown et al, 2020).…”

“…Prediction has been proposed as an organizing principle of human cognition in general and language in particular (Dell et al, 2021;Pickering and Garrod, 2013). In machine learning, deep-learning language models trained to predict upcoming words-or, more generally, some aspect of their input from another ("self-supervised learning")-have been immensely successful as a foundation for language technologies (Peters et al, 2018;Devlin et al, 2019), and have been shown to learn a surprising amount about language structure (Linzen and Baroni, 2021).…”

Repetition parallels in language and motor action: Evidence from tongue twisters and finger fumblers.

“…Second, the pattern of onset errors in the manual task—showing local effects from one position to the next but no effects of the global sequence pattern—brought to our attention a crucial difference between the tasks’ stimuli: nonwords in the speech task were monosyllabic, and each nonword appeared on the screen as a single unit, with no clear divide between the onset and the offset (see also Dell et al, 2021; Fischer-Baum et al, 2021, for a discussion of the tight organization of the syllable unit). In the manual task, however, onset and offset components were clearly separated on the left versus right sides of the screen and were performed with separate effectors (left hand, right hand).…”

Language generality in phonological encoding: Moving beyond Indo‐European languages

“…Virtually all models implement an online process of syllabification that assigns selected segments to syllabic roles, as in Dell's ( 1986 ) well-known use of a syllable frame to syllabify segments ( see also Sevald et al, 1995;Shattuck-Hufnagel, 1979;Smolensky et al, 2014 ). Syllabification in this case both orders segments within a syllable and provides the basis for ( learned ) constraints on phonological encoding, such as phonotactic constraints ( Alderete & Tupper, 2018;Dell et al, 2000Dell et al, , 2021Goldrick & Daland, 2009;Vousden et al, 2000 ). In some models, syllables are represented as chunks and play a role of mediating the activation flow between a word and its component sounds.…”