Category: Talks

Compositionality and Generalization in Emergent Languages

Rahma Chaabouni, Facebook AI Research

Tuesday, Nov 24 2020 11:00-12:00 GMT
Zoom Details: [Please Request]

Natural language allows us to refer to novel composite concepts by combining expressions denoting their parts according to systematic rules, a property known as compositionality. Linguists agree that compositionality plays a crucial role in natural language, accounting for its productivity (i.e., the ability to express an infinite number of ideas by finite means).
Despite the importance of productivity, deep neural networks, known for their impressive NLP performances, often fail to generalize to unseen examples. In this talk, I ask whether the language emerging in deep agents possesses a similar ability to refer to novel primitive combinations and whether it accomplishes this by strategies akin to intuitive compositionality. I address this by first looking at the generalization performances of deep agents on unseen inputs. Our experiments show that given sufficiently large input spaces, the emergent language will naturally develop the ability to refer to novel composite concepts. Second, I introduce different measures of intuitive compositionality that the community is looking at. Based on these measures, I show experiments demonstrating no correlation between the degree of compositionality of an emergent language and its ability to generalize. I conclude the talk by suggesting that, while compositionality is not necessary for generalization, it provides an advantage in terms of language transmission: The more compositional a language is, the more easily it will be picked up by new learners.

Bias in the Learning of Sound Patterns: An Experimental Investigation

Eleanor Glewwe, Grinnell University

Tuesday, 17.11.2020
16:00 – 17:00
Room: [virtual Zoom talk]

A persistent question in the field of phonology is why the phonological typology exhibits asymmetries. For instance, word-final devoicing (/rad/ → [rat]) is common cross-linguistically, but word-final voicing (/rat/ → [rad]) is not observed. One hypothesis for why these asymmetries exist is that language learners are biased against acquiring certain types of sound patterns. One way of testing for such learning biases is by using artificial grammar learning experiments. Artificial grammar learning studies investigating biases in phonological learning have uncovered robust evidence for complexity bias (a bias against featurally complex patterns) but little for naturalness bias (a bias against phonetically unnatural patterns) (Moreton & Pater 2012). I present two phonotactic learning experiments that tested for both complexity bias and naturalness bias by comparing how well participants learned different distributions of a stop voicing contrast (/t/ vs. /d/). Together, the two experiments offer mixed evidence for naturalness bias but stronger evidence for complexity bias, while also demonstrating how the broader phonological structure of an artificial language affects performance. Based on the results of my experiments and a review of the experimental literature, I argue that we must distinguish between perceptually-rooted and articulatorily-rooted naturalness bias and claim that only perceptual naturalness, not articulatory naturalness, biases phonological learning.

Mechanisms underlying universals in vocal communication patterns in songbirds

Jon Sakata, Department of Biology, McGill University

Tuesday, 10.11.2020
16:00 – 17:00
Room: [virtual Zoom talk]

The structure and patterning of many learned behaviors are more similar across populations than expected by chance. Such “universals” in learned behaviors are prevalent not only in humans but also in non-human animals, and interestingly, the structures of some universals are highly conserved across humans and non-human animals. Given the prevalence of such behavioral commonalities, it is important to reveal the mechanisms that contribute to the emergence of shared structures and patterns. My lab is investigating mechanisms underlying universals in learned communicative behaviors in songbirds. In this talk, I will discuss our recent experiments highlighting the contributions of biological predispositions in learning and motor production biases to the emergence of universal patterns in vocal communication in songbirds. In addition, I will highlight the various parallels in universal vocal patterns between humans and songbirds and the potential role of shared sensorimotor processes in generating these common patterns.

Mechanisms of conversation: Audience design and memory

Si On Yoon, University of Iowa

Tuesday, 03.11.2020
16:00 – 17:00
Room: [virtual Zoom talk]

Communicating with others is one of the most fundamental social activities of everyday life. Even though communication plays an undeniably important role in our lives, the mechanisms of language processing used in conversation are largely unexplored due to the difficulties in examining natural conversational language with traditional psycholinguistic approaches. In a newly developed experimental paradigm to study conversation in the lab, I have been able to examine how speakers tailor language during multiparty conversation (one speaker and two listeners). I have also expanded this paradigm to further look at how speakers balance the needs of the different partners for successful communication in conversations with up to 7 people. In another line of work, to examine the nature of the memory representations that are built and used during natural communication, I test individuals with hippocampal amnesia and severe memory impairment, as well as healthy older adults. These results demonstrate the extent to which tailoring language to one’s audience requires intact hippocampal-dependent memory systems. Across these lines of inquiry, my research is uncovering critical aspects of the cognitive mechanisms that afford, not just language production and comprehension, but more specifically communication through natural conversation.

Competition, selection and communicative need in language change: an investigation using corpora, computational modelling and experimentation

Andres Karjus, University of Edinburgh

Thursday, 29.10.2020
13:30 – 14:00
Room: [virtual Zoom talk]

Constant change is one of the few truly universal cross-linguistic properties of living languages. In this thesis I focus on lexical change, and ask why the introduction and spread of some words leads to competition and eventual extinction of words with similar functions, while in other cases semantically similar words are able to companionably co-exist for decades.

I start out by using extensive computational simulations to evaluate a recently published method for differentiating selection and drift in language change. While I conclude this particular method still requires improvement to be reliably applicable to historical corpus data, my findings suggest that the approach in general, when properly evaluated, could have considerable future potential for better understanding the interplay of drift, selection and therefore competition in language change.

In a series of corpus studies, I argue that the communicative needs of speakers play a significant role in how languages change, as they continue to be moulded to meet the needs of linguistic communities. I developed and evaluated computational methods for inferring a number of linguistic processes — changes in communicative need, competition between lexical items, and changes in colexification — directly from diachronic corpus data. Applying these new methods to massive historical corpora of multiple languages spanning several centuries, I show that communicative need modulates the outcome of competition between lexical items, and the colexification of concepts in semantic subspaces.

I also conducted an experiment in the form of a dyadic artificial language communication game, the results of which demonstrate how speakers adapt their lexicons to the communicative needs of the situation. This combination of methods allows me to link actions of individual speakers at short timescales to population-level findings in large corpora at historical timescales, in order to show that language change is driven by communicative need.

Learnability and cross-linguistic constraints on the meanings of clause-embedding predicates

Wataru Uegaki, University of Edinburgh

Tuesday, 27.10.2020
11:00 – 12:00
Room: [virtual Zoom talk]

A central question in semantic research is whether there are any cross-linguistically robust constraints on the possible denotations of lexical items of certain grammatical categories. Recent work has explored cross-linguistic constraints in the domain of clause-embedding predicates like “know”, “agree”, and “wonder” (Spector & Egré 2015; Theiler, Roelofsen & Aloni 2018; Uegaki 2019; Steinert-Threlkeld 2020). Within this line of work, two basic questions can be distinguished. The first is empirical: Which constraints, if any, do we find in the semantics of clause-embedding predicates? The second is theoretical: What may explain the existence of such universal semantic constraints? In this talk, I will talk about my recent collaborative projects on both of these fronts. Specifically, I will discuss two empirical proposals concerning cross-linguistic constraints on the meanings of clause-embedding predicates (joint work with Floris Roelofsen), and report on artificial-language experiments that aim to tap into how these constraints relate to learnability (joint work with Jennifer Culbertson).

Learning verb second (in the lab)

Marc Meisezahl, University of Edinburgh

Tuesday, 13.10.2020
11:00 – 12:00
Room: [virtual Zoom talk]

Verb second (V2) is a word order pattern attested only in a small number of languages. In this type of construction, the finite verb strictly follows the first constituent regardless of its grammatical function or category. Considered diachronically, V2 has been stable in some languages while others have lost it (Holmberg 2015). In the light of findings showing the impact of learning biases on typological patterns (Culbertson, Smolensky & Legendre 2012), the question is raised why not all languages have lost V2. More specifically, what properties of the language are relevant for the retention and hence the acquisition of V2. A potentially important factor is the frequency of the different types of preverbal elements. In this talk, I present the results of an artificial language learning experiment in which participants were exposed to a V2 language with different types of constituents that are equally likely to appear preverbally. The goal of this experiment is twofold: First and more basically, it is assessed whether V2 can be learned in an experimental setting. Secondly, the experiment establishes a baseline for future experiments with non-uniform distributions of preverbal elements. Furthermore, I will present ideas for a corpus study in which the frequencies of different types of preverbal constituents are determined. These results may be used for follow-up experiments on the role of preverbal constituents.

How Translatable are Common Words? Some Answers from Distributional Semantic

Bill Thompson Princeton University

Tuesday, 06.10.2020
16:00 – 17:00
Room: [virtual Zoom talk]

We analysed the semantic networks of 1,016 concepts in 41 languages using distributional models of lexical semantics. We examined which semantic domains (e.g. animals, emotions, body parts and numbers) show the most and least alignment between different languages, and whether alignment is greater for more concrete terms (it is not). We examined how alignment varies for different parts of speech, and how it relates to human judgements of similarity and to lexical factors such as frequency and neighborhood density. Our analyses show that the alignment between one language and another is statistically related to the cultural and historical relatedness of the languages, offering large-scale statistical evidence for the view that natural language lexical semantics are influenced by processes of cultural evolution.

Rerent properties and word order in emerging communication systems

Fiona Kirton The University of Edinburgh

Wednesday, 30.09.2020
15:00 – 15:30
Room: [virtual Zoom talk]

Why do languages look the way they do? This question lies at the core of much of linguistics research, and answering it can shine a light on the relationship between individual cognitive biases and linguistic structure. One area that has attracted particular attention is basic word order. Many languages exhibit a fixed or dominant ordering of subject (S), object (O), and verb (V); it has been suggested that SOV is the natural ordering of entities in an event, and the default order used by all newly emerging languages. A growing body of research has investigated this question using the silent gesture paradigm in which participants describe events using only gesture and no speech. This work has uncovered a range of factors that influence the way people convey information about events in the absence of linguistic conventions, challenging the view that there is a single natural order. One particularly fruitful line of work has focused on the relationship between animacy and word order, and a number of hypotheses have emerged that seek to explain the nature of this relationship.

In this thesis, I present a number of experiments that build on this work to investigate how properties of individual referents influence word order choices in emerging communication systems. Using a range of experimental approaches, I investigated three proposed drivers of word order variation: salience, modality, and communication. The results of this work support the hypothesis that the salience of referents is a key factor in shaping word order in emerging communication systems. However, the findings were inconclusive about the precise mechanisms underlying this relationship. I suggest that further work is required to better understand the contribution of factors such as modality and native language interference in experimental settings. I also highlight the need for further research to understand how language producers negotiate the communicative challenge of accurately conveying information about events in which the role of referents is ambiguous.

The forms and meanings of grammatical markers support efficient communication

Frank Mollica The University of Edinburgh

Tuesday, 29.09.2020
11:00 – 12:30
Room: [virtual Zoom talk]

Functionalist accounts of language suggest that forms are paired with meanings in ways that support efficient communication. Previous work on grammatical marking suggests that word forms have lengths that enable efficient production, and previous work on the semantic typology of the lexicon suggests that word meanings represent efficient partitions of semantic space. Here we present an integrated information-theoretic framework that captures how communicative pressures influence both form and meaning. We apply the framework to the grammatical features of number, tense, and evidentiality, and show that in all three cases the framework explains both which systems of feature values are attested across languages and the relative lengths of the forms for those feature values.