What is human language, how is it learned, and how did it evolve? We invite people who are interested in these questions to talk about their research. 

Katrin Schulz, Raquel Alhama, Fausto Carcassi, Marieke Schouwstra

** sign up for our email list to receive talk announcements: **

https://groups.google.com/g/lelams 

speaker: Tessa Verhoef (LIACS, Leiden)

title: The emergence of language universals in neural agents and vision-and-language models

location: PCH 6.05. We will offer this as a hybrid event; meeting Meeting ID: 878 2270 6729

abstract:

Human cognition constrains how we communicate. Our cognitive biases and preferences interact with the processes that drive language emergence and change in non-trivial ways. A powerful method to discern the roles of cognitive biases and processes like language learning and use in shaping linguistic structure is to build agent-based models. Recent advances in computational linguistics and deep learning sparked a renewed interest in such simulations, creating the opportunity to model increasingly realistic phenomena. These models simulate emergent communication, referring to the spontaneous development of a communication system through repeated interactions between individual neural network agents. However, a crucial challenge in this line of work is that such artificial learners still often behave differently from human learners. Directly inspired by human artificial language learning studies, we proposed a novel framework for simulating language learning and change, which allows agents to first learn an artificial language and then use it to communicate, with the aim of studying the emergence of specific linguistics properties. I will present two studies using this framework to simulate the emergence of a well-known language phenomenon: the word-order/case-marking trade-off. I will also share some very recent findings where we test for the presence of a well-known human cross-modal mapping preference (the bouba-kiki effect) in vision-and-language models. Cross-modal associations play an essential role in human language understanding, learning, and evolution, but our findings reveal that current multimodal language models do not align well with such human preferences.

speaker: Andres Karjus (Tallinn University and Estonian Business School)

title: Scaling the scientist: large language models in language and text research.

location: PCH6.31. We will offer this as a hybrid event; meeting Meeting ID: 820 9076 9012 

abstract: 

The increasing capacities of instructable multimodal large language models (LLMs) have presented an unprecedented opportunity to scale up data analytics in sciences dealing with language, text and visual data, and to automate qualitative tasks previously typically allocated to human labor. Of particular interest to the humanities and social sciences is the capacity to use them as zero-shot classifiers and inference engines. While classifying texts or images for various properties has been long available in the form of supervised learning, the necessity to train (or tune pretrained) models on sufficiently large sets of labeled examples complicates their adoption in research beyond generic tasks like sentiment analysis, where prepackaged solutions are often available. Approaches like word or sentence embeddings and topic modeling allow for explorative approaches but are typically laborious to interpret and difficult to use for confirmatory inference. This talk discusses recent research on using LLMs in zero-shot classification and reasoning scenarios, their feasibility as replacement for typical distant reading toolsets, and potential pitfalls. LLM outputs naturally contain errors (as does human annotation), but the error rate can and should be included in subsequent statistical modeling. A bootstrapping approach is discussed, which in turn can be easily integrated in a quantitizing mixed methods research design, advocated for here as one particularly fit for purpose framework for leveraging machine scalability while supporting replicability and transparency. 

speaker: Polina Tsvilodub (University of Tübingen)

title: How to be relevantly overinformative to a polar question: Reasoning about questioner goals to provide more relevant answers

location: PCH 4.11. We will offer this as a hybrid event; meeting Meeting ID: 869 9939 5665

abstract: 

Imagine you are working as a barista at a coffeeshop. A customer asks a polar question like “Do you have iced tea?” but you’ve run out. In this situation, you might likely provide an  overinformative answer going beyond a simple “yes” or “no” (e.g., “No, but we’ve got iced coffee!”), but what principles guide the selection of additional information?

This talk proposes that such answers draw on learning about our interlocutors from language; they present a non-trivial instance of pragmatic communication which depends for complex reasoning drawing on these inferences about the interlocutors and world knowledge. Specifically, I will argue that respondents use the uttered question in order to reason about the questioner’s preferences and goals, and craft answers relevant to those goals (Hawkins & Goodman, 2017). I will provide experimental evidence from several human studies suggesting that overinformativeness in human answering is driven by considerations of relevance to the questioner’s goals which respondents flexibly adjust given the functional context in which the question is uttered, and given the shared world knowledge. Furthermore, I will present a Rational Speech Act model (RSA, Frank & Goodman, 2012) of pragmatic overinformative question answering which builds on an action-based notion of relevance of information. It captures qualitative patterns in speakers’ utterance choices across a variety of contexts.

Finally, I will present a comparison of these human and probabilistic modeling results to question-answering performance of a variety of state-of-the-art neural language models. We find that most models fail to adjust their answering behavior in a human-like way and tend to include irrelevant information (Tsvilodub et al., 2023). We show that most recent fine-tuned LLMs are highly sensitive to the form of the prompt and only achieve human-like answer patterns when guided by an example of a relevantly overinformative answer or a cognitively-motivated explanation of the complex reasoning.

speaker: Marieke Woensdregt (RU Nijmegen)

title: The role of social cognition and social interaction in shaping language

location: PCH 4.04. We will offer this as a hybrid event; meeting Meeting ID: 869 9939 5665

abstract: 

In this talk, I will discuss the roles that social cognition (specifically: perspective-taking) and social interaction (specifically: asking for clarification) play in shaping language. On the role of social cognition, I will present experimental work in which we manipulated the possibility/difficulty of perspective-taking in a task where pairs of participants had to develop a novel communication system. Preliminary analyses suggest that when participants have access to each other’s perspectives (i.e., when the partner’s avatar is visible) they converge on a deictic system, which is also associated with higher communication accuracy. When the partner’s avatar is invisible, however, deictic systems break down and participants gravitate towards object labelling. These results suggest that the communicative context influences the emergence and functioning of different communicative systems through perspective-taking. On the role of social interaction, I will present computational modelling work that investigates how noise and other-initiated repair (i.e., asking for clarification) interact, through cultural evolution, with the structure of language. Earlier work has shown that compositional structure can arise under the combined pressures of (i) learnability and (ii) expressivity. Here we connect these results to two other ubiquitous features of human communication: noise and interactive repair. We find that even in the absence of a learnability pressure, compositional languages are favoured somewhat under the combined pressures of (i) mutual understanding and (ii) minimal effort, because compositional structure allows for efficient other-initiated repair.

speaker: Dan Dediu (UB Barcelona)

title: Linguistic diversity: from weak individual biases to large-scale structural differences between languages

location: PCH 5.37. We will offer this as a hybrid event; meeting ID: 842 6155 5885. 

abstract: It has become a platitude, but it is still worth repeating that the 7000 or so language currently used differ in many ways at basically every level ones cares to look at, but that this variation is patterned and resulting from a multitude of complex processes. In this talk I will try to argue that some of this variation is due to the amplification of (relatively weak) biases expressed at the level of the individual language learner and user, as language is repeatedly used and transmitted within and between groups and generations. The engine driving this amplification is represented by cultural evolutionary processes acting in structured communicative networks. I will present some examples of such such amplification arguably having taken place, with the amplified biases ranging from phonetic details due to minute variation in vocal tract anatomy, to the colour lexicon due to environmental effects on the physiology of the eye, passing through the complex origins of such biases in the interactions between our genes, environment and culture. Finally, I will briefly present some ongoing modelling and experimental work that tries to understand how communication in structured networks might explain this type of amplification.