Conceptual Structure Is Shaped By Competing Pressures For Simplicity And Informativeness
Jon Carr (University of Edinburgh)
Tuesday 10 April 2018, 11:00–12:30
G32 7 George Square
Languages are shaped by competing pressures from learning and communication. Learning favours simple languages, while communication favours informative ones, giving rise to the simplicity–informativeness tradeoff. Languages that evolve under this tradeoff are both maximally simple (learnable) and maximally informative (communicatively useful). This has been shown in natural language and in experimental settings. For example, Kemp and Regier (2012) showed that kinship systems exist at the optimal frontier of simplicity and informativeness. In a separate line of experimental work, Kirby, Tamariz, Cornish, and Smith (2015) showed that when artificial languages evolve under a learning pressure alone, they become simple and uninformative, and when languages evolve during communication, they become complex and informative; it is only when both pressures are at play that we find languages at the optimal frontier.
However, a recent iterated learning experiment by Carstensen, Xu, Smith, and Regier (2015) showed that artificial languages expressing spatial relationships tended to become more informative when subjected to a pressure from learning. This is a surprising result given the previous work briefly reviewed above, which says that informativeness is driven by the pressure from communication, not from learning. One potential explanation for this result lies in their measure of informativeness, communicative cost, which is sensitive to (a) the number of words that the language is comprised of (expressivity) and (b) the extent to which similar meanings are expressed by the same word (which we will term convexity). In their experiment expressivity was fixed at four words. As a result, the reduction in communicative cost they found must be due to categories evolving to become more convex, i.e. picking out increasingly tightly-clustered sets of meanings.
To demonstrate that learning favours convex categories, we conducted two experiments in which participants learned and produced a category system for stimuli varying on two dimensions, size and angle. In Experiment 1 participants were trained on one of three systems: One marking a distinction in angle, one marking a distinction in size, and one marking a distinction on both dimensions (see Fig. 1). The results indicated that the Angle-only system was easiest to learn, followed by the Size-only system; the Angle & Size system was hardest to learn, despite having the lowest communicative cost.
In Experiment 2, the output of one participant became the input to the following participant in a standard iterated learning design. An example chain is shown in Fig. 2. Over 12 generations the category system became increasingly easy to learn, as indicated by decreasing intergenerational transmission error. Furthermore, in the majority of chains, the language converged on a system marking only a distinction on the angle dimension, which participants found easiest in Experiment 1. This increase in simplicity is driven by increasing convexity.
We also found that most chains converged on fewer than four categories. This suggests that iterated learning acts as a pressure for simplicity by simultaneously decreasing expressivity and increasing convexity. However, if, as in Carstensen et al. (2015), expressivity is held constant, the learning pressure can only act through convexity: Although languages may become more informative under iterated learning, they do so not because of a pressure to be more communicatively useful, which in Carstensen et al.’s study necessarily decreases communicative cost as a side-effect of increasing convexity. This therefore suggests that, contra Carstensen et al. (2015), languages which are both simple yet informative will only emerge when pressures from learning and communication are at play. We support these conclusions with a Bayesian iterated learning model that displays strikingly similar results.
A rational model of linguistic accommodation and its potential role in language simplification
Stella Frank (University of Edinburgh)
Tuesday 10 April 2018, 11:00–12:30
G32 7 George Square
Languages with large numbers of adult learners tend to be less morphosyntactically complex than languages where adult learners are rare (Wray & Grace, attributed to deficiencies in adult language learning. Here we investigate an additional or alternative mechanism: rational accommodation by native speakers to non-native interlocutors.
Humans have a general aptitude for reasoning about the knowledge, beliefs and motivations of other individuals, including their linguistic knowledge (e.g. Clark, 1996; Ferguson, 1981). While our interlocutors’ linguistic knowledge will often be close to our own, this may not be the case in a population with many non- native speakers. We introduce a rational model of interactions between individuals capable of reasoning about the linguistic knowledge of others, and investigate the case of a non-native speaker interacting with an native speaker who reasons about their linguistic knowledge and accommodates accordingly. Our model shows that this accommodation mechanism can lead to the non-native speaker acquiring a language variant that is less complex than the original language.
We assume a simple model in which a language consists of a distribution over linguistic variants (e.g. past tense forms). Language simplification is modelled as regularisation, whereby the most frequent variant becomes more frequent; this corresponds to, and can be measured as, entropy reduction. We model the interaction between a non-native speaker and a native speaker as interaction between two rational (Bayesian) agents. Both agents have the same initial priors and update their beliefs about the language from data in the same way, but the non-native speaker has simply seen much less data. Within an interaction, the native speaker has a parametrisable tendency to accommodate to the non-native speaker: instead of simply using their own language, they use the version of the language that they believe the non-native speaker may have acquired at this stage of their learning, given limited exposure. Importantly, the native speaker does not know exactly what data the non-native has seen. Instead, the native speaker models the non- native speaker’s linguistic knowledge by integrating over possible datasets the non-native speaker might have seen.
Representative model results for a sample language are shown in Figure 1. While learners interacting with non-accommodating speakers eventually learn the original language, non-native speakers interacting with accommodating native speakers end up learning a more regular language. This is due to the combination of the limited exposure of the non-native individual, which results in highly skewed initial distributions and some probability of not having seen low-frequency variants (Hertwig, Barron, Weber, & Erev, 2004; Hahn, 2014), in conjunction with a native speaker who is aware of and accommodates this initial bias in the non- native speaker’s input, therefore providing the non-native speaker with further data which ‘locks in’ their biased starting point.
This model shows that accommodation by native speakers to non-native speakers during interaction can lead to language simplification, and therefore suggests how accommodation can explain the link between population makeup and lin- guistic complexity. The model assumes that individuals are capable of reasoning rationally about their interlocutors’ linguistic knowledge, an assumption we are currently testing empirically with human learners.