Same But Different: Systems for Smoothing Noun Entropy in Communication in German and English

Language is a defining human characteristic, and languages define cultures. But how does language actually work? Traditional answers to this question embrace two key assumptions: compositionality, which holds that the meanings of messages are built from semantic 'atoms' that are modified by syntactic rules; and transfer, the idea that speakers' signals encode meanings that are extracted by listeners. While these assumptions accord well with our intuitions, trying to formalise them has led to claims that languages themselves are unlearnable and that much linguistic knowledge is innate (a hypothesis which we suggest is fatally undermined by recent developments in machine learning), as well as engendering a perspective that considers many complex systems within languages - such as the German gender system - to be functionless ornaments. Our proposed research program will directly challenge this last assumption. Our research proposal explicitly rejects compositionality and assumes that - rather than facilitating transfer of meanings - human communication serves to reduce uncertainty about intended meaning. Within this framework, our previous work has shown that while the child's linguistic environment may be impoverished for learning 'languages as traditionally conceived', its statistical properties abound with information that can allow them to master the communication processes defined by information theory. Moreover, the ways in which children's learning mechanisms develop makes them ideally receptive to this statistical input. Our proposal extends our approach to grammatical gender, an aspect of language that has defied traditional 'intuitive' attempts to explain its function (and perhaps because of this, is currently a candidate for linguistic re-engineering in many communities). Our key claim is that under an information theoretic approach - where communication is a mutually predictive process - the role of grammatical gender becomes clear: it is a solution to the challenge of supporting the processing of nouns which are the least predictable part of language. By comparing German and English - gendered and non-gendered languages - we will show how grammatical gender systems play a critical role in reducing the communicative uncertainty associated with nouns. In doing so, we aim to show how all languages are structured to solve this problem, and establish what these structures and patterns of usage can tell us about the nature of linguistic processes. Specifically, the project has three aims: To use corpus analysis to explore how grammatical gender systems help balance (smooth) information in noun phrases. Investigate the learnability of these systems and the factors influencing this using artificial language learning experiments. To experimentally elicit spontaneous speech to: (i) see how and when native speakers use the structures we identify; (ii) investigate speakers' sensitivity to these factors; (iii) examine their effects on communicative behaviour in real time. In doing this, we aim to provide opinion/decision makers with a clearer understanding of the costs and benefits of gendered languages. More broadly, in showing how humans learn communicative codes, and by modelling their properties, we seek to overcome traditional stumbling blocks that obscure our understanding of language and put forward a coherent theory of how it actually works.