Collins Conference Room
Seminar
  US Mountain Time

Our campus is closed to the public for this event.

Zipf's law of abbreviation and the principle of least effort
Jasmeen Kanwal (University of Edinburgh)

As Zipf observed in 1935, human languages (and possibly also animal communication systems) universally exhibit an inverse relationship between word length and word frequency; the higher the frequency of a word, the shorter it tends to be. Zipf hypothesised that this universal ‘Law of Abbreviation’ (ZLA) arises from speakers operating under a Principle of Least Effort (PLE). When combined with a pressure to communicate successfully, the pressure to minimise effort for speakers would lead to an optimal solution wherein the shortest forms are mapped to the most frequent meanings. However, an alternative hypothesis states that ZLA can be explained purely by invoking cognition-external statistical facts about randomly generated systems (e.g. Moscoso del Prado, 2013). If Zipf’s hypothesis is correct, then we should observe speakers actively optimising form-meaning mappings in line with ZLA in a simple communication task. We investigate this using an artificial language communication game administered on Amazon Mechanical Turk. Our results support Zipf’s hypothesis, and emphasise the broader view that, to understand the origin of universal structural properties of language such as ZLA, we must zoom in to the micro-evolutionary scale and investigate how the behaviour of individual speakers might drive large-scale evolution towards universal patterns.

Neutral models of language change
Kevin Stadler (University of Edinburgh)

Much current research into the pressures that shape the evolution of languages over time emphasizes the adaptive nature and consequent selection of linguistic traits. In this talk I will make a case for neutral models of language change: I will argue that individual instances of language change should not be construed as replicator selection, i.e. that we should not assume any inherent asymmetry between competing linguistic variants to be responsible for their selection. I will give a general overview and critique of both neutral and (replicator-)selection models of language change, and show how emergent trend-amplifying biases can offer a better explanation for the sporadic and arbitrary nature of language change events. In doing so I will also point to the next challenges in understanding the dynamics of linguistic evolution, namely the complex interactions between the generation of (linguistic) variants and their consequent spread through large speech communities.

Learners regularize homonyms and synonyms similarly
Vanessa Ferdinand (Santa Fe Institute)

An important topic in language evolution is the emergence of linguistic structure from interacting individuals. Individual learners also impose structure on language by eliminating unpredictable variation in a variety of learning tasks (e.g. Hudson Kam & Newport, 2005; Reali & Griffiths, 2009; Smith & Wonnacott, 2010; Perfors, 2012; Ferdinand, Thompson, Kirby, & Smith, 2013). This elimination of variation is known as regularization. The human “regularization bias” is often assumed to be a monolithic bias that operates similarly across all forms of linguistic variation. However, different kinds of linguistic variation play different roles in communicative success and it is reasonable to assume that learners regularize such variation differently. For example, synonymous variation (when one meaning has two signals) allows the accurate recovery of a meaning whereas homonymous variation (when one signal has two meanings) does not. I present an experiment that compares the regularization of synonymous and homonymous mappings in an artificial language learning task and show that participants regularize both mappings similarly, but not identically: participants appear to encode synonymous variation accurately, but homonymous variation is encoded as more regular. Additionally, I explain how linguistic variation can be quantified with various information-theoretic metrics and show how these metrics outperform the established approach of identifying regularization by the raw frequencies of linguistic variants.

More SFI Events