Search

Self-paced reading has been a widely used experimental method for study of the processing of sentences and texts. In this paper, we introduce the method to the Czech audience. We summarize its advantages and limitations and provide practical suggestions on stimuli construction and data pro-cessing. We also present different variants of the method, we discuss its ecological validity, and we summarize the experimental evidence showing that reaction times collected in self-paced reading can be linked to processing demands people might experience during reading. Finally, we present three examples of the application of the method: an experiment on agreement attraction in Czech, an experiment on garden-path sentences in Czech and an experiment studying the processing of short discourses in English. We also briefly discuss new trends that connect corpus linguistics with psycholinguistic discourse processing research and lead to the development of reading-time corpora.

This study compares 16 vowel-normalization methods for purposes of sociophonetic research. Most of the previous work in this domain has focused on the performance of normalization methods on steady-state vowels. By contrast, this study explicitly considers dynamic formant trajectories, using generalized additive models to model these nonlinearly. Normalization methods were compared using a hand-corrected dataset from the Flemish-Dutch Teacher Corpus, which contains 160 speakers from 8 geographical regions, who spoke regionally accented versions of Netherlandic/Flemish Standard Dutch. Normalization performance was assessed by comparing the methods' abilities to remove anatomical variation, retain vowel distinctions, and explain variation in the normalized F0–F3. In addition, it was established whether normalization competes with by-speaker random effects or supplements it, by comparing how much between-speaker variance remained to be apportioned to random effects after normalization. The results partly reproduce the good performance of Lobanov, Gerstman, and Nearey 1 found earlier and generally favor log-mean and centroid methods. However, newer methods achieve higher effect sizes (i.e., explain more variance) at only marginally worse performances. Random effects were found to be equally useful before and after normalization, showing that they complement it. The findings are interpreted in light of the way that the different methods handle formant dynamics.

This paper explores the possibility that the spread of sound change within a community correlates with individual differences in imitation capacities. The devoicing of labiodental fricatives in Dutch serves as a case study of an ongoing sound change showing regional and individual variation. The imitation capacities of Dutch speakers born and raised in five regions of the Dutch language area were investigated in a forced imitation task (Study 2) and a spontaneous imitation task (Study 3), and compared to baseline productions (Study 1) of the variable undergoing sound change. Results showed that the leaders of sound change in each region were significantly less accurate in imitating model talkers – when they were instructed to – than conservative speakers, but they were more inclined to spontaneously imitate talkers. These insights are discussed in view of the literature on different types and measures of imitation capacities, on the actors of sound change and the two apparently paradoxical features of the language system: its stability and its potential for sound change.

The expression of reflexivity in Indonesian and related languages is based on various strategies, see Cole and Hermon (2005), Kartono (2013), and Schadler (2014) for discussion. This paper focuses on the expression of reflexivity based on the element diri and its cognates, not discussed in these papers. As a reflexive marker, bare diri is not specified for grammatical features such as number, gender, and person, so it imposes no restrictions on the value of the subject argument. It is only allowed with a subset of verbs, namely agent–theme verbs. Our goal is to determine its precise role. After applying a number of diagnostics for argumenthood (Dimitriadis and Everaert 2014), we show that diri is not an argument. We propose that the role of diri is that of an element marking detransitivization of the verb and reflecting an operation combining the latter's agent and theme roles into one complex agent–theme role ("a bundling operation" in the sense of Reinhart and Siloni 2005). This complex role is assigned to the remaining argument resulting in a reflexive interpretation. Further tests also show that agent and patient roles are indeed present in verbs with diri after the bundling operation.

Modified numerals, such as at least three and more than five, are known to sometimes give rise to ignorance inferences. However, there is disagreement in the literature regarding the nature of these inferences, their context dependence, and differences between at least and more than. We present a series of experiments which sheds new light on these issues. Our results show that (a) the ignorance inferences of at least are more robust than those of more than, (b) the presence and strength of the ignorance inferences triggered by both at least and more than depends on the question under discussion (QUD), and (c) whether ignorance inferences are detected in a given experimental setting depends partly on the task that participants are asked to perform (e.g., an acceptability task versus an inference task). We offer an Optimality Theoretic account of these findings. In particular, the task effect is captured by assuming that in performing an acceptability task, participants take the speaker’s perspective in order to determine whether an expression is optimal given a certain epistemic state, while in performing an inference task they take the addressee’s perspective in order to determine what the most likely epistemic state of the speaker is given a certain expression. To execute the latter task in a fully rational manner, participants have to perform higher-order reasoning about alternative expressions the speaker could have used. Under the assumption that participants do not always perform such higher-order reasoning but also often resort to so-called unidirectional optimization, the task effect finds a natural explanation. This also allows us to relate our finding to asymmetries between comprehension and production that have been found in language acquisition.

Production-based cognitive models, such as Adaptive Control of Thought-Rational (ACT-R) or Soar agents, have been a popular tool in cognitive science to model sequential decision processes. While the models have been useful in articulating assumptions and predictions of various theories, they unfortunately require a significant amount of hand coding, both with respect to what building blocks cognitive processes should consist of and with respect to how these building blocks are selected and ordered in a sequential decision process. Hand coding of large, realistic models poses a challenge for modelers, and also makes it unclear whether the models can be learned and are thus cognitively plausible. The learnability issue is probably most starkly present in cognitive models of linguistic skills, since linguistic skills involve richly structured representations and highly complex rules. We investigate how reinforcement learning (RL) methods can be used to solve the production selection and production ordering problem in ACT-R. We focus on four algorithms from the (Formula presented.) learning family, tabular (Formula presented.) and three versions of deep (Formula presented.) networks (DQNs), as well as the ACT-R utility learning algorithm, which provides a baseline for the (Formula presented.) algorithms. We compare the performance of these five algorithms in a range of lexical decision (LD) tasks framed as sequential decision problems. We observe that, unlike the ACT-R baseline, the (Formula presented.) agents learn even the more complex LD tasks fairly well. However, tabular (Formula presented.) and DQNs show a trade-off between speed of learning, applicability to more complex tasks, and how noisy the learned rules are. This indicates that the ACT-R subsymbolic system for procedural memory could be improved by incorporating more insights from RL approaches, particularly the function-approximation-based ones, which learn and generalize effectively in complex, more realistic tasks.

In this paper, we introduce a novel domain-general, statistical learning model for P&P grammars: the Expectation Driven Parameter Learner (EDPL). We show that the EDPL provides a mathematically principled solution to the Credit Problem (Dresher 1999). We present the first systematic tests of the EDPL and an existing and closely related model, the Naïve Parameter Learner (NPL), on a full stress typology, the one generated by Dresher & Kaye’s (1990) stress parameter framework. This framework has figured prominently in the debate about the necessity of domain-specific mechanisms for learning of parametric stress. The essential difference between the two learning models is that the EDPL incorporates a mechanism that directly tackles the Credit Problem, while the NPL does not. We find that the NPL fails to cope with the ambiguity of this stress system both in terms of learning success and data complexity, while the EDPL performs well on both metrics. Based on these results, we argue that probabilistic inference provides a viable domain-general approach to parametric stress learning, but only when learning involves an inferential process that directly addresses the Credit Problem. We also present in-depth analyses of the learning outcomes, showing how learning outcomes depend crucially on the structural ambiguities posited by a particular phonological theory, and how these learning difficulties correspond to typological gaps.

This study investigates the link between the perception and production in sound change in progress, both at the regional and the individual level. Two devoicing processes showing regional variation in Dutch are studied: the devoicing of initial labiodental fricatives and of initial bilabial stops. Five regions were selected, to represent different stages of change in progress. For each region, 20 participants took part in production (Study 1) and perception (Study 2) experiments. First, the results of the production tasks give additional insight in the regional and individual patterns of sound change. Second, the regional perceptual patterns in fricatives match the differences in production: perception is the most categorical in regions where the devoicing process is starting, and the least categorical in regions where the process of devoicing is almost completed. Finally, a clear link is observed between the production and perception systems undergoing sound change at the individual level. Changes in the perceptual system seem to precede changes in production. However, at the sound change completion, perception lags behind: individuals still perceive a contrast they no longer produce.

This paper reports on an explorative sociophonetic study of the phoneme /s/ in the Dutch language area. Our aim is to investigate the regional variation in the realisation of this phoneme, and to test experimentally the observation of Collins & Mees (2003) that /s/ is sometimes pronounced more like [ ʃ ], especially in the Randstad area (called s-retraction). One hundred native speakers of Dutch produced nineteen monosyllabic words containing /s/ in different syllabic contexts. The speakers were born and raised in one of five regions of the Dutch language area (West Flanders, Flemish Brabant, Netherlands Limburg, South Holland and Groningen). Spectral centre of gravity (CoG) and duration were used to measure the degree of s-retraction. CoG values turned out to be significantly lower (consistent with more retraction) in the regions in The Netherlands than in the Flemish regions. Speakers from South Holland produced significantly shorter /s/ than the other speakers. In conclusion, /s/ shows patterns of regional variation that are not fully in line with the observation forwarded by Collins & Mees (2003). The difference between the Flemish and Dutch regions shows that s-retraction is found in an area larger than the Randstad, possibly pointing towards a North-South pattern of variation.

AM dependency parsing is a method for neural semantic graph parsing that exploits the principle of compositionality. While AM dependency parsers have been shown to be fast and accurate across several graphbanks, they require explicit annotations of the compositional tree structures for training. In the past, these were obtained using complex graphbank-specific heuristics written by experts. Here we show how they can instead be trained directly on the graphs with a neural latent-variable model, drastically reducing the amount and complexity of manual heuristics. We demonstrate that our model picks up on several linguistic phenomena on its own and achieves comparable accuracy to supervised training, greatly facilitating the use of AM dependency parsing for new sembanks., Accepted at the 5th Workshop on Structured Prediction for NLP (http://structuredprediction.github.io/SPNLP21)

This paper explores how the rational theory of memory summarized in Anderson (1991) can inform the computational psycholinguistic models of human parsing. It is shown that transition-based parsing is particularly suitable to be combined with Anderson's theory of memory systems. The combination of the rational theory of memory with the transition-based parsers results in a model of sentence processing that is data-driven and can be embedded in the cognitive architecture Adaptive Control of Thought-Rational (ACT-R). The predictions of the parser are tested against qualitative data (garden-path sentences) and a self-paced reading corpus (the Natural Stories corpus).

Phonological processes tend to be defined over natural classes (Chomsky & Halle 1968), but there are some arbitrary and language-specific aspects to class behaviour (e.g., Mielke 2004). This paper shows that it is possible to implement a procedure of finding language specific natural classes using contrast detection (Dresher 2014, Sanstedt 2018), but in standard OT with domain-general methods. Three toy languages are constructed, based on those in Prickett & Jarosz (2021), in which /e/ raises to [i] in the presence of a high vowel and in which /s/ palatalizes to [ʃ] before [i]. In one language, raising feeds palatalization (transparent); in the second, raising counterfeeds palatalization (opaque); in the third, raising applies transparently, but only in certain morphemes (lexically specific). All three languages are learned with a version of Round’s (2017) learner that learns indexed constraints (Pater 2000) that are attached to specific segments in morphemes rather than entire morphemes (cf. Nazarov 2021). This learner is able to find appropriate natural classes for these data, both phonetic natural classes (=traditional natural classes) and what I call phonologically natural classes (classes defined by having certain phonetic properties and undergoing a range of phonological processes), showing the feasibility of this approach.

In this article we report on a survey that was conducted on Bonaire, one of the six Caribbean islands that were formerly known as the Netherlands Antilles. The majority of the Bonairean population speaks Papiamentu, a Spanish/Portuguese lexifier creole, as their mother tongue. Dutch used to be the only officially recognized language on the island up until 2007, when the Netherlands Antilles recognized Papiamentu and English as co-official languages besides Dutch. Since the dissolution of the Netherlands Antilles on October 10 of 2010, Bonaire has adopted a new political status as an exceptional municipality of the European Netherlands. This political reform has a strong impact on the small island community due to demographic growth, the influence of European Dutch legislation and the increasing cost of living (Bak-Piard 2010). The status of Papiamentu has changed dramatically as it is no longer recognized as an official language, its use as a home language is in decline and its role in the education system is under attack (Bak-Piard 2016). This article presents the results of 262 questionnaires that were distributed on Bonaire to investigate patterns of language use, language attitudes and identity among the Bonairean population. The results pointed out that the speech community of Bonaire is rather homogeneous. Papiamentu is widely used and attitudes toward Papiamentu and Bonairean identity are overall (very) positive. These findings provide insights that are important for the development of a sustainable language policy in the education system, serving Bonairean students to achieve their full potential.

Seemingly simple sentence like ‘Sue had three husbands’ have a reading that cannot be explained by standard scope mechanisms such as quantifier raising (Szabó 2011; Francez 2018). On this so-called summative reading, ‘three husbands’ is neither de re nor de dicto w.r.t. the past tense operator. Rather, the numeral appears to take independent scope: there are three individuals such that for each of these it is the case that there was a time in the past such that person became Sue’s husband. Solutions to this puzzle tend to stipulate an interpretation of the numeral that allows it to take exceptional scope using familiar devices like quantifier raising (e.g. Francez 2018). Here I explore the value of an entirely different route. Given the fact that numerals can be adjective-like, summative readings could be an instance of exceptional adjectival scope (Abusch 1997). That is, the phenomenon could be parallel to the weird scope-taking of ‘unknown’ in sentences like ‘The suspect stayed in an unknown hotel’. The resulting analysis will not be very far removed from non-restrictive or post-suppositional analyses of numerals that have been proposed in the literature (e.g. Solt 2009; Brasoveanu 2013). This analysis has the advantage of being able to account for readings that are simultaneously summative and cumulative. However, I also show how all proposals tend to predict that summative readings are much more common than they in fact are.

Multilingual language models exhibit better performance for some languages than for others (Singh et al., 2019), and many languages do not seem to benefit from multilingual sharing at all, presumably as a result of poor multilingual segmentation (Pyysal o et al., 2020). This work explores the idea of learning multilingual language models based on clustering of monolingual segments. We show significant improvements over standard multilingual segmentation and training across nine languages on a question answering task, both in a small model regime and for a model of the size of BERT-base.

Can language models learn grounded representations from text distribution alone? This question is both central and recurrent in natural language processing; authors generally agree that grounding requires more than textual distribution. We propose to experimentally test this claim: if any two words have different meanings and yet cannot be distinguished from distribution alone, then grounding is out of the reach of text-based models. To that end, we present early work on an online game for the collection of human judgments on the distributional similarity of word pairs in five languages. We further report early results of our data collection campaign.