One of the key features of natural languages is that they exhibit long-distance filler-gap dependencies (FGDs): In the sentence `What do you think the pilot sent __?' the wh-filler what is interpreted as the object of the verb sent across multiple words. The ability to establish FGDs is thought to require hierarchical syntactic structure. However, recent research suggests that recurrent neural networks (RNNs) without specific hierarchical bias can learn complex generalizations about wh-questions in English from raw text data (Wilcox et al. 2018; 2019). Across two experiments, we probe the generality of this result by testing whether a long short-term memory (LSTM) RNN model can learn basic generalizations about FGDs in Norwegian. Testing Norwegian allows us to assess whether previous results were due to distributional statistics of the English input or whether models can extract similar generalizations in languages with different syntactic distributions. We also test the model's performance on two different FGDs: wh-questions and relative clauses, allowing us to determine if the model learns abstract generalizations about FGDs that extend beyond a single construction type. Results from Experiment 1 suggest that the model expects fillers to be paired with gaps and that this expectation generalizes across different syntactic positions. Results from Experiment 2 suggest that the model's expectations are largely unaffected by the increased linear distance between the filler and the gap. Our findings provide support for the conclusion that LSTM RNN's ability to learn basic generalizations about FGDs is robust across dependency type and language.

LSTMs Can Learn Basic Wh- and Relative Clause Dependencies in Norwegian

We explore pupil diameter (PD) as estimator of Sense of Embodiment (SoE) using data of three user studies. We hypothesize that pupil diameter reflects SoE in a direct and indirect way. If individuals feel strongly embodied, presenting an emotional stimulus like a threat to the surrogate will produce a strong response, as if the stimulus would be presented to their own body. This would lead to a positive correlation between SoE and pupil dilation during the presentation of emotional stimuli. Besides this direct effect, there may also be an indirect effect. It is postulated that higher degrees of embodiment reduce workload when controlling a surrogate. This indirect effect of embodiment through lower workload on the PD would result in a negative correlation between SoE and PD since lower workload results in smaller PD. These direct and indirect effects were partially confirmed by the results of three experiments. We observed that PD and SoE are positive and direct correlated in case of emotional stimuli subjected to the surrogate (e.g. a threat), and that PD tended to be smaller for participants who experienced a condition designed to provide high SoE compared to one designed to provide low SoE.

Pupil Diameter as Implicit Measure to Estimate Sense of Embodiment

While an apparent tendency for negative markers to appear before the verb has been observed in typology, language acquisition, and language emergence, it remains uncertain what factors may motivate such a preference. The present study uses an artificial language learning paradigm to test the existence of learning asymmetries consistent with Neg-First preferences in English speakers. The study further incorporates a dyadic interaction task to investigate proposals that the Neg-First tendency is driven by communicative factors. Results show that learners overall produced more preverbal negation than was found in the input language, consistent with a Neg-First bias. However, interaction only induced greater preverbal negation use when preverbal negation was the majority word order in the input language. This does not support the proposal that communication generally promotes a Neg-First bias, but is suggestive of greater regularization when a production bias is aligned with a bias to eliminate variability during communication.

Investigating the Presence of NegFirst Biases in Learning and Communication

Syntactic harmony occurs when heads and dependents align within and across different types of phrases in a language. Harmony is a well-known (statistical) typological universal: in most languages, many if not all heads and dependents are consistently ordered (i.e., either head-dependent, or dependent-head). Despite decades of work, from every conceivable theoretical perspective, the origins of syntactic harmony remain opaque. However, recent work using artificial language learning has suggested that harmonic patterns are easier to learn than their non-harmonic counter-parts. Thus at least part of the explanation for this tendency may be linked to learning. Here, we explore whether the mechanism behind the learning bias for syntactic harmony is fundamentally domain-general by instantiating harmony in non-linguistic stimuli. Our findings support the claim that the origins of syntactic harmony lie in a domain-general bias for simplicity acting on linearized, language-specific categories.


Syntactic harmony arises from a domain-general learning bias

Research has shown that comprehenders represent alternative meanings to single focused words online, consistent with Rooth’s (1992) formal semantic account. However, focus can also take a broader scope over whole phrases ('read the manuscript'). We examined whether in these cases, alternatives are represented by testing for an interference effect of the particle only. Using probe recognition, we first tested unmentioned alternatives to the constituent parts of VPs, to object nouns (Exp 1, 'letter' for 'manuscript') and verbs (Exp 2, 'wrote' for 'read'). In Exp 3, we tested alternatives to whole phrases ('wrote the letter'). In all experiments, alternative probes were processed slower than unrelated ones. We found varying evidence of the interference effects of only with noun, verb and whole-phrase alternatives. Overall, this study does not provide support for the generalisation of the effects of only to broad focus. Additionally, we discuss the methodological implications of probing whole phrases.


The Comprehension of Broad Focus: Probing Roothian Alternatives

Agreement attraction (i.e. facilitatory interference manifested by sped-up reading times) observed in establishing subject-verb number agreement by comprehenders when reading ungrammatical sentences with number-matching attractor nouns, has been long-established and cross-linguistically validated. For languages with rich inflectional morphology, case syncretism has been suggested to play a role in the phenomenon. In the present self-paced reading study on Czech, we show that unlike in other languages, facilitatory interference is not observed and that not even case syncretism is sufficient for its appearance. We put forward several possible explanations for this anomaly exhibited by Czech compared to other languages. We propose that the lack of semantic agreement in the language could be one of these. Finally, we discuss the implications of these results for the models of long-distance dependency resolution in comprehension.

No agreement attraction facilitation observed in Czech: Not even syncretism helps

Fast and efficient correction of speech errors is essential to effective communication. Yet, despite several accounts of error detection, no computational account exists to explain how humans repair their speech errors. This paper proposes the first such model. We demonstrate that a simple automatic mechanism can form the basis of most repairs. We then demonstrate that augmenting the model with a conflict-based monitoring-control loop allows it to capture more nuanced findings in human speech error repair data.

A conflict-based model of speech error repairs in humans

Listeners interpret spoken language differently, depending on the perceived identity of the speaker. Within increasingly diverse societies, it is important to understand how speaker characteristics such as accent may impact simple linguistic exchanges. In the current study, children are being tested on their ability to recall factual information provided by native and non-native accented English speakers. Preliminary results demonstrate that children tend to remember neutral facts better if they heard them from a native accented speaker. Interestingly, children understood the native and non-native accented talkers equally well, indicating that comprehension difficulties could not account for the results. To date, there have been no studies examining how a speaker’s accent affects children’s memory, and only preliminary, mixed results in adults have been found. This gap in the literature underscores the need to understand how linguistic biases might impact information transfer across diverse language communities.

Where's the Bingleduff? How speaker accent shapes children's linguistic processing

Listeners struggle to identify talkers with a different accent than their own, a phenomenon known as the Other Accent Effect (OAE). But for reasons that are not well understood, the OAE is not consistently observed in all studies. Comprehension-related processing demands offer one explanation, such that other-accented talkers who are more easily understood are also easier to recognize. Here, we test this hypothesis using a forensic-style voice line-up. We examine native English-speaking adults’ ability to recognize talkers from four accent groups, manipulating comprehension-related processing demands by presenting listeners with predictable sentences and subtitles (low-demand condition), or variable sentences without subtitles (high-demand condition). As predicted, the OAE was only observed for talkers with non-native accents. But crucially, our comprehension manipulation had no impact on talker recognition accuracy of any accent type. We conclude that comprehension ease is likely not a key factor driving the OAE. Other possible explanations are discussed.

Do I need to repeat myself? Getting to the root of the Other Accent Effect

An important step in children's socio-cognitive development is learning how to engage in coordinated conversations. This requires not only becoming competent speakers but also active listeners. This paper studies children's use of backchannel signaling (e.g., "yeah!" or a head nod) when in the listener's role during conversations with their caregivers via video call. While previous work had found backchannel to be still immature in middle childhood (i.e., 6 to 11 years of age), our use of both more natural/spontaneous conversational settings and more adequate controls allowed us to reveal that school-age children are strikingly close to adult-level mastery in many measures of backchanneling. The broader impact of this paper is to highlight the crucial role of social context in evaluating children's conversational abilities.

Backchannel Behavior in Child-Caregiver Zoom-Mediated Conversations

While the grammatical use of space for referential strategies is attested across many sign languages of Western Europe, Kata Kolok, a rural sign language used in Northern Bali, has not developed anaphoric pointing in space nor agreement verbs (Engberg-Pedersen, 1993; Liddell, 2003; de Vos, 2012). To find out whether such typological differences can be explained by differences in the respective co-speech gesture systems, this preregistered study is collecting data for a cross-cultural comparison. Building on Motamedi et al. (2021), we are conducting studies in Bali and the Netherlands using an iterated learning silent gesture paradigm in which hearing people communicate transitive events using only gestures. Preliminary data indeed suggest that our Balinese participants do not employ space the same way our Dutch participants do. We will present comparative analyses and evaluate the role of co-speech gesture systems in sign language emergence in the lab or when evolving from spontaneous interaction.

Cross-cultural differences in the emergence of referential strategies in artificial sign languages

Each grammatical construction has its own function, and typically multiple constructions are combined to express a message. When the functions of two constructions conflict in a way that cannot be reconciled, their combination is judged ungrammatical. Here we consider one such type of case: “syntactic island violations.” Specifically, we consider combinations of wh-questions with 11 other constructions. Wh-questions request direct information about a particular constituent. Using a new Discourse task, we quantify how directly 11 constructions convey information in simple declarative sentences. Results demonstrate acceptability judgments on the wh-questions correlate with the degree to which the 11 constructions convey information directly. Thus, we argue that degrees of unacceptability of “island violations” result from the extent to which the discourse functions of the constructions involved conflict (N=240).

Underline digital video library

LSTMs Can Learn Basic Wh- and Relative Clause Dependencies in Norwegian

Pupil Diameter as Implicit Measure to Estimate Sense of Embodiment

Investigating the Presence of NegFirst Biases in Learning and Communication

Syntactic harmony arises from a domain-general learning bias

The Comprehension of Broad Focus: Probing Roothian Alternatives

No agreement attraction facilitation observed in Czech: Not even syncretism helps

A conflict-based model of speech error repairs in humans

Where's the Bingleduff? How speaker accent shapes children's linguistic processing

Do I need to repeat myself? Getting to the root of the Other Accent Effect

Backchannel Behavior in Child-Caregiver Zoom-Mediated Conversations

Cross-cultural differences in the emergence of referential strategies in artificial sign languages

Islands effects without extraction: the discourse functions of constructions predicts island status

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES