A Minimalist Model and Computational Toolkit for Code Switching

Abstract

Researchers have long noted the fact that multilingual speakers often engage in code switching: That is, they use mixes of multiple languages within single utterances or discourses. Many previous accounts of the phenomenon have treated the separate languages involved as distinct from one another, and there have been a number of code switching-specific principles and rules proposed that aim to "constrain the interactions of the two [grammatical] systems in mixture" (MacSwan 2000). The converse assumption — i.e., that there is a single set of grammatical principles governing both code-switched and monolingual utterances — has gained traction recently, and the Minimalist Program provides an appealing framework for exploring this kind of approach because it explicitly posits that the computational system for language is invariant across languages, with the lexicon being the sole locus of cross-linguistic variation. By examining the interaction between multiple lexicons and the competing demands that each places simultaneously upon a universal computational system, the Minimalist analysis presented in this work sheds light on the nature of the elements and operations involved in syntactic computations, even in single-language contexts. In particular, an investigation of English-Chinese code switching data from Singaporean speakers suggests that there is a wide range of potential interactions between lexical items from different languages within a single sentence, across both their underlying structures and surface forms. This provides general support for a realisational view of the derivational process (as in the Distributed Morphology framework) and motivates a fundamental distinction between the phenomena of code switching (whereby the features and forms of successive items come from different languages), calquing (whereby the features of an item come from one language, but its surface form has been translated into another), and borrowing (whereby the features and surface form of an item have their origin in one language, but the item has been adapted fully into the lexicon of another). In addition, this work also introduces the Code Switching Toolkit (CSTK), a computational modelling platform that implements a fundamental set of Minimalist principles while remaining extensible enough to accommodate internal differences between individual theories that come under the Minimalist umbrella. The CSTK allows researchers to model syntactic derivations, visualise the hierarchical structures that are built up along the way, see exactly when and why each of these derivations eventually crash or converge, and refine their accounts accordingly. This is not only useful for theoretical exploration, but also has potential utility in the generation of synthetic code-switched data for NLP and other machine learning applications.