Language Evolution and Computation

Tuesday 26th, 11-12.30, B21 7 George Square

Motivated vs. conventional systematicity: Implications for language learning and the structure of the lexicon

Alan Nielsen

Given that the task of language learning involves pressures for both learnability and communicative accuracy, one might expect that, contra to the linguistic dogma of arbitrariness, systematic mappings between signals and meanings might be easier to learn. In the past decade two streams of research — one exploring systematic signal-meaning mappings that are conventional, and one exploring mappings that are sound-symbolic or otherwise motivated by the structure of the language learner/environment. In this talk I will present the results of an experiment where participants learned a language that used either motivated (vowel height/frontness–> size; consonant plosivity –> shape) or conventional associations between signals and meanings. The results of this experiment suggest that sound-symbolic associations are easier to learn than conventional associations, and that at least part of this benefit arises from an early advantage for motivated associations. However, the results also serve to demonstrate how little is known about what consonantal and vowel characteristics drive sound-symbolic associations.

Tuesday 19th November, 11-12.30, B21 7 George Square

The elimination of unpredictable variation is dependent on speaker identity

Olga Feher

Natural languages do not normally exhibit unpredictable variation, which is when an object is labelled by two or more words that are used interchangeably in a random manner. Synonyms are common but their use is usually conditioned on some social or contextual variable. Therefore, unpredictable variation is a good tool to study language evolution and learners’ biases. We trained participants on a miniature artificial language using a word learning task to test the effects of speaker identity on the elimination of unpredictable variation. The spoken input language contained two labels for each object presented, one label used twice as often as the other but the variation was distributed differently across 1-3 speakers. There were four conditions: in one, participants heard a single speaker name all objects; in the second condition 3 speakers named objects with variable usage; in the third condition 3 speakers used the labels categorically: two speakers using one label all the time and the third speaker using the other label consistently; and in the last condition is like the third with the odd label always used by the same “odd ball” speaker. We found that participants who received variable input from all the speakers tended to probability match, whereas when variation came from between speakers, participants were much more likely to regularise in their word recall. This can be due to conformity effects or learning the categorical behaviour of speakers.

Time: 11-12.30, Tuesday 12th November
Place: B21 7 George Square

Semantic structure and emerging conventions in silent gesture

Marieke Schouwstra

When individuals do not share a common language, they cannot reliably use existing linguistic conventions when they communicate. This is is the case in, e.g., unsupervised second language acquisition and home sign. I will briefly review observations from these systems, showing that they are largely governed by semantic organisational principles (Jackendoff, 2002).

Subsequently, I will present results from my silent gesture experiments, in which naive participants were asked to describe events using only gesture and no speech (similarly to Goldin-Meadow et al., 2008). I will show that, like unsupervised second language learners and homesigners, silent gesturers shape utterances flexibly and according to their meaning:

– Whereas for simple events that involve motion through space (e.g., ‘pirate smashes guitar’), SOV basic word order is preferred, more abstract intensional events (e.g., ‘pirate searches guitar’) give rise to a different ordering: SVO.

– When participants are asked to describe events that take place at some other time than now (e.g., ‘a pirate smashes a guitar at three o’clock’), they systematically apply a strategy similar to that observed in adult second language learners and homesigners, in which the temporal information precedes—and never interrupts—the event information (such as in ‘3 O’CLOCK-PIRATE-GUITAR-THROW’). This mirrors the semantic representation of a temporally displaced event.

These observations strengthen the hypothesis that in language systems without full syntax, meaning determines structure, suggesting an evolutionary scenario in which semantic structure preceded syntactic structure. However, how and why have these early semantic stages of language evolved such that the relationship between utterance structure and semantic structure became obscured? In other words, how did language go from semantic organising principles to fully functioning syntactic rules? I will conclude with some ideas how to investigate this question empirically, using various silent gesture lab experiments that incorporate repeated social transmission.

Tuesday 29th October, 11.00-12.30, B21 7 George Square

Klaas Seinhorst – The learnability of phoneme inventories

I will present my PhD project “The learnability of phoneme inventories”, which I started at the University of Amsterdam in September 2012. The starting point of my project was the observation that languages have strong preferences for certain sets of plosive sounds, while other sets are extremely rare. In the experiments I plan to conduct, human participants will learn sets of linguistic stimuli analogous to sound systems – both frequent and infrequent ones; using iterated learning, I will monitor the evolution of these sets and compare the outcomes to typological data. If the experimental findings comply with the typological results, this would point to an important role of learning biases in phonological typology.

Tuesday 22nd October, 11-12.30, B21 7 George Square

Model Fitting and Prediction for Language Evolution: Quirks and Opportunities
Bill Thompson

Questions concerning the origins of linguistic structures are increasingly being studied by means of artificial language learning experiments. Central to this enterprise is the drive to uncover cognitive biases that shape the evolution of linguistic systems. A particularly promising growth area in this field is the use of cognitive model fitting techniques. In this talk I’ll argue that, perhaps unlike comparable topics in cognitive science, questions of language evolution must be seen in the light of cultural evolution. This feature at once poses potentially unique difficulties and exciting opportunities in experimental design and model fitting. I’ll consider several techniques that may help maximise confidence in the inferences we can draw, from experimental data, about language learning biases and their consequences for population-level linguistic phenomena. In particular I hope to show that the quirks of studying a culturally learned behaviour, and the opportunities to make predictions about culture by harnessing information in experimental learning data, are served equally by the same approach.

Bill has unfortunately had to postpone his talk next week, so I will be stepping in to give a short presentation about some of the work that Kenny and I have been doing on analysing iterated artificial sign language experiments. It’s really fun and rich data, and we’d like your input on how to tackle digging through what we seem to have found!

—-

Title: The cultural evolution of sign languages in the laboratory: from pantomime to linguistic system?

Speaker: Simon Kirby

Place: B21, 7 George Square

Date: 1st October, 11.00-12.30

—-

See you there, I hope!

Simon

Tue 24th September, 11.00-12.30, B21 7 George Square

Experimental and theoretical models of cultural evolution

Marius Kempe, University of Durham

Abstract:
I will talk about two experimental and two theoretical models of cultural evolution. In the first experiment, I test the hypothesis that increasing group size speeds up cultural accumulation, using a novel puzzle-solving task and within a transmission chain design. I find support for this hypothesis, in contrast with previous experiments. In the second experiment, also using a transmission chain design, I examine perceptual errors in recreating Acheulean handaxes and ask whether such errors can account for the variability of Acheulean technology over time. Using the accumulated copying error model to compare the experimental data to archaeological records, I conclude that perceptual errors alone were likely not the driving force behind Acheulean evolution. In the first theoretical study, I present models of cultural differences between populations and of cumulative culture, which build on existing models and accord with empirical data. I then show that the models, when combined, have two qualitative regimes which may correspond to human and nonhuman culture. In the second theoretical study, I present a ‘fundamental theorem of cultural selection’, an parallel to Fisher’s Fundamental Theorem of Natural Selection for cultural evolution. I discuss how this theorem formalizes and sheds light on cultural evolutionary theory.

We’re starting the new semester of LEC talks with a short talk from Cem Bozsahin, an old friend of the LEC who is passing through Edinburgh. This will be a dry run for a talk at the Philosophy and Theory of AI conference, so we can expect the meeting to take substantially less than an hour even with discussion.

NOTE: WE WILL BE IN ROOM B21, 7 GEORGE SQUARE

Tuesday, 17th September, 11 am.

Cem Bozsahin
Cognitive Science Department,
METU Informatics Institute, Ankara

Natural Recursion Doesn’t Work that Way

All hierarchicaly organized observed behaviors are instances of recursion by value. Recursion by name can be shown to be more powerful than recursion by value: the former has infinite types, the latter does not.

Any animal that can plan has recursion, so that that kind of recursion is probably not unique to humans. Humans appear to have a certain kind of recursion which is unique, the kind that works with embedded push-down automaton, and that is probably not unique to language.

Therefore it is unhelpful to build entire conception of recursion in natural language and humans on a much powerful notion of recursion than needed, and on purely syntactic terms rather than conceptual or semantic, including theories about its evolution.