ABSTRACTS

Abstract

Frequency plays a key role in language dynamics in domains spanning learning, memory, and communication. For example, frequency correlates with morphological irregularity (high frequency words are more likely to be irregularly inflected) and high frequency units are also more likely to be regularised by learners (i.e., exhibit less intra- and inter-speaker variation). In this paper, I present an agent-based model which looks at the tension between these two patterns. Intuition might suggest these dynamics would lead to systems which gradually eliminate variation, but this is clearly not the pattern we find in natural languages (e.g., Cuskley et al., 2014; Cuskley et al., 2017). Using a variant of the regularity game (Cuskley et al., 2018), this model reframes questions of variation in terms of how language systems maintain frequency dependent variation, exploring the relative contributions of maturation (i.e., that child learners have different biases than adult learners), memory, and socio-communicative pressures. In particular, the tension between learning/memory and communication supports dynamics similar to stabilizing selection in biology: for highly frequent forms, the model shows frequency dependent selection against morphologically regular (i.e., type-dominant) rules where this would disrupt communicative success in populations. The work concludes with an echo of calls to reorient thinking about language variation: in many more traditional approaches to linguistics, the draw of universalist explanations still tacitly frames variation as a phenomena which must be explained away in order to reveal "core" properties of human language. Instead, I argue we should refocus on variation itself as a core adaptive feature of language, shifting to precise questions regarding how language maintains variation as fodder for selection, and thus, adaptation.

Abstract

The study of general mechanisms of language change tends to focus on either universal cultural selective pressures, such as learnability, expressivity or ease of production/perception, or on extralinguistic conditions such as speaker population size or proportions of second language speakers. However, factors intrinsic to the linguistic system in place may also play an important role in constraining or triggering change. Languages are built up of structural traits, forming the grammar of the language. Grammatical traits exhibit empirical dependencies, that have traditionally been considered universal. Recently the focus has shifted to scrutinizing lineage-specific characteristics, suggesting that language evolution is partially guided by local dynamics leading to distinct basins of attraction. Our aim here is to explore whether these basins can be identified through the examination of interactions between grammatical traits at a global level. We use Grambank, a new database of 195 traits for 2,467 language varieties, to create a network of linguistic traits, where the weight of the edge between two traits is determined by their likelihood of co-occurence in languages. The network is weighted to correct for common ancestry between languages, using scores of lexical proximity from A Global Lexical Database (GLED). Preliminary results from community detection in the network indicate attraction between similar patterns concerning bound or free morphology respectively and word order, that go beyond language families.

Abstract

Language contact refers to a situation of two languages which are spoken in a given society. The main question is then about dominance, extinction or coexistence of the two languages. Following the milestone paper by Abrams and Strogatz [1] two main parameters, prestige and volatility, appear in different models considering these questions. I will address the role of covert prestige in terms of different linguistic preferences for the two speech communities. I will also consider the role of heterogeneous volatility, either as Heterogeneous Language volatility or as Heterogeneous Social volatility. Given the many possible outcomes depending on parameter values, I will argue that we might need data about these parameters themselves, rather than fitting their values to reproduce language use data with a specific model.

[1] D. M. Abrams and S. H. Strogatz, “Modelling the dynamics of language death”, Nature 424, 900–900 (2003).

Abstract

The aim of this paper is to propose and assess metrics that can be used for the structural evaluation of text of a historical language generated by AI. More specifically, the paper focuses on the generation of Old English, of which the total of written records is slightly over 3 million words (Taylor et al. 2003; Healey et al. 2004) and clearly insuJicient for AI models. The procedure of text generation is prompting with an AI chat (Petroni et al. 2021) , which is inputted a poetic text of Old English and is asked to produce a prose version of the poetry. Zero-shot learning and few-shot learning (with examples of prosification of other texts) are carried out (Gao et al. 2021). The output is assessed as to quantity and as to quality. On the side of quantity, the ratio of words and the ratio of unique words of the new version are calculated. As for quality, the meaning and the structure of the text are taken into account. Meaning, including lexical attestedness and collocation, is given more weight (60%) than structure, which includes inflection and word order and has a 40% weight. An Overall Acceptability Score is put forward that constitutes a weighted average of four diJerent scores: Attestedness Score, Inflection Score, Collocation Score and Word Order Score. Scaling factors adjust for the fact that the diJerent scores make reference to diJerent units (words, pairs and sentences). The resulting Overall Acceptability Score is checked against an analysis of the constructions in the generated prose. Conclusions are expected regarding the weights of the scores.

References

• Gao, Tianyu, Adam Fischz & Danqi Cheny. 2021. Making Pre-trained Language Models Better Few-shot Learners. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 3816–3830.

• Healey, A. (ed.), J. Price-Wilkin & X. Xiang. 2004. The Dictionary of Old English web corpus. Toronto: Dictionary of Old English Project, Centre for Medieval Studies, University of Toronto.

• Petroni, F., T. Rocktäschel, P. Lewis, A. Bakhtin, Y. Wu, A. H. Miller & S. Riedel. 2019. Language Models as Knowledge Bases? Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 2463–2473.

• Taylor, A., A. Warner, S. Pintzuk & F. Beths 2003. The York-Toronto-Helsinki Parsed Corpus of Old English Prose [https://www-users.york.ac.uk/~lang22/YcoeHome1.htm].