Competition and selection in the multilingual mind : a dynamic approach to language change

Abstract

Languages change because of variation and the way in which features are selected and replicated. In heterogeneous multilingual societies where widespread bilingualism is the norm, it is expected that language contact affects the outcome of this selection process. Building on work by Croft (2000), Mufwene (2001, 2008), Ansaldo (2009), and Aboh (2015), it is assumed that language change at the individual level results from contact between different idiolects, and that features of those idiolects with similar functions compete for selection. Research on contact-induced language change suggests that feature selection is constrained by factors such as frequency, salience, semantic content, congruence, and transparency (Aboh and Ansaldo, 2007; Siemund and Kintana, 2008). However, these constraints are hard to quantify in a principled manner, especially when considered in a multilingual context. This dissertation investigates how and to what extent it is possible to computionally quantify the properties of features in competition, to help interpret the process of competition and selection, and ideally, predict the outcome of language contact.

Functional load, defined as the relative contribution of an individual morpheme to the overall sentence given a multilingual feature pool, is proposed as a constraint on selection. Three multilingual recursive neural networks trained on learning abstract bilingual sentence representations of (Mandarin) Chinese and English are employed (Le and Zuidema, 2014; Hermann and Blunsom, 2014) that incorporate principles of distributional and compositional semantics. That is, the models use frequency-based co-occurrence statistics and information on the compositional structure of sentences to learn word and sentence representations in both languages. Models that employ these properties have shown success at capturing meanings of words and phrases (Erk, 2012) and can enable us to quantify the functional load of individual morphemes cross-linguistically. The three models differ with respect to either the tasks on which they are trained or the input they receive and all make slightly different predictions with respect to functional load. The overall hypothesis that holds for all models states that when a speaker has the option to select features from different systems, a feature with a higher functional load will more likely get selected.

The results from the models suggest that the functional load hypothesis can account for several patterns found in the contact variety Colloquial Singapore English (CSE). For example, one of the distinctive features of this mixed variety is the widespread use of sentence-final particles (Wong, 2004). Sentence-final particles get assigned a relatively high functional load in the models, and accordingly, make up strong competitors and are therefore more likely to get selected. Additionally, the functional load of several English morphemes significantly correlated with the omission rate of these morphemes in CSE. The findings suggest that the quantified notion of functional load can indeed be posited as a constraint on selection and can inform studies on contact-induced language change both at the individual and societal level. Moreover, it shows that computational models are not only useful as functional or analytical tools but can be used to directly inform linguistic research.