Search

Language is acquired within a complex, interactiveenvironment. A key question for cognitive science is whetherand how different types of environmental cues might affectthe learning and processing of language. In this paper, weexplore the role of feedback as a possible cue in a novelactive artificial language learning task: The Picture GuessingGame. Subjects were instructed to guess which scenecorrectly displayed the meaning of a spoken sequence ofunfamiliar monosyllabic words. After their response, eitherpositive, negative, or no feedback was provided. Theprediction was that feedback would help the subject toeventually learn the vocabulary, syntax, and semantics of theartificial language. The results indeed showed that feedback(both positive and negative) is beneficial and necessary toattain a certain level of learning. Interestingly, the datashowed that positive feedback may be particularly helpful forthe learner, promoting more in-depth learning of the artificiallanguage.

Language exhibits striking systematicity in its form-meaningmappings: Similar meanings are assigned similar forms. Herewe study how systematicity relates to another well-studiedphenomenon, linguistic regularization, the removal of unpre-dictable variation in linguistic variants. Systematicity is ulti-mately a property of classes of form-meaning mappings, eachmember of which can be acted upon independently by linguis-tic regularization. Both are supported by a cognitive bias forsimplicity, but this leaves open the question of how they inter-act to structure the lexicon. Using data from a recent artificialgesture learning experiment by Verhoef, Padden, and Kirby(2016), we formalize cognitive biases at the item level and thelanguage level as inductive biases in a hierarchical Bayesianmodel. Simulated data from models that lack either one ofthose biases show how both are necessary to capture subjects’systematicity preferences. Our results bring conceptual clar-ity about the relationship between regularization and system-aticity and promote a multi-level approach to cognitive biasesin artificial language learning and language evolution.

Speakers tend to assign more linguistic material to less predictable elements. This tendency is typically explained by a bias for an efficient trade-off between production effort and understandability and is claimed to shape linguistic structures across languages. Recent work suggests this trade-off enters the linguistic system through learning processes with learners deviating from their input by increasing marking for less predictable elements. However, no study to date has tested whether child learners also show such predictability-based marking, an important gap seeing that children are the primary learners in real-life language acquisition. A recent study showed that adults increase predictability-based marking of an optional-plural marker, in line with communicative efficiency. Here, we ask if children show a similar pattern. Results show that children, unlike adults, do not show an efficient trade-off in their productions. We discuss implications for the role of different language learners on language change.

One of the central goals of language evolution research is toexplain how systematic structure emerges. A culturalevolutionary approach proposes that the systematic structure oflanguage arises from the use and transmission of language.Motamedi and colleagues (2016) investigated the influences ofthese forces on the evolution of language by generating anartificial sign language in the lab. Over several generations ofnew learners and their interactions, an initially unsystematic setof silent gestures developed markers for functional categoriesof person, location, object, and action. Here we describe resultsof two studies that compared the learnability of solo-producedpantomimes versus signals that had been transmitted and usedby interlocutors. In these studies, participants saw an artificialsign and judged whether an English translation matched ormismatched the meaning of the sign. In an event-relatedpotential (ERP) study, we found that mismatches elicited largernegativities in the ERP than matches. However, those effectswere most reminiscent of the classic N400 response in theevolved signs. This study provides a clearer view on how themechanisms that drive language evolution change language toadapt to a learner’s brain.

Research on the acquisition of morphology commonly predicts that agglutinating systems should be easier to learn than fusional systems. This is argued to be due to compositional transparency: the mapping between morphemes and meanings is one-to-one in agglutinating systems, but not in fusional systems. This is supported by findings in first and second language learning (Goldschneider & DeKeyser 2001, Slobin 1973), typology (Dressler 2003, Haspelmath & Michaelis 2017), and language evolution (Brighton 2002). We present findings from a series of artificial language learning experiments which complicate this picture. First, we show that when only two features (e.g., NOUN CLASS and NUMBER) are morphologically encoded, the learnability of fusional and agglutinating systems does not differ significantly. This finding holds when learners are given an additional cue to morpheme segmentation–which in principle should make the agglutinating system easier. However, the error patterns of the two groups provide some evidence that learners might have a bias for transparent structures. Our results suggest that the advantages of agglutinating over fusional systems may be overstated, particularly when a small number of features are encoded. Since agglutinating systems likely bear additional costs (e.g., segmentation, longer word length, and the online cost of mapping between morphemes and meanings), such systems do not guarantee learning ease under all circumstances.

A foundational goal of linguistics is to investigate whethershared features of the human cognitive system can explainhow linguistic patterns are distributed across languages. Inthis study we report a series of artificial language learning ex-periments to test a hypothesised link between cognition and apersistent regularity of morpheme order: number morphemes(e.g., plural markers) tend to be ordered closer to noun stemsthan case morphemes (e.g., accusative markers) (Greenberg,1963). We argue that this typological tendency may be drivenby a bias favouring orders that reflect scopal relationships inmorphosyntactic composition (Bybee, 1985; Rice, 2000; Cul-bertson & Adger, 2014). We taught participants an artificiallanguage with noun stems, and case and number morphemes.Crucially, the input language indicated only that each mor-pheme preceded or followed the noun stem. Examples inwhich two (overt) morphemes co-occurred were held out—i.e.,no instances of plural accusatives. At test, participants wereasked to produce utterances, including the held-out examples.As predicted, learners consistently produced number closer tothe noun stem than case. We replicate this effect with freeand bound morphemes, pre- or post-nominal placement, andwith English and Japanese speakers. However, we also findthat this tendency can be reversed when the form of the casemarker is conditioned on the noun, suggesting an influence ofdependency length. Our results provide evidence that univer-sal features of cognition may play a causal role in shaping therelative order of morphemes.

It is often assumed that cross-linguistically more prevalent distinctions are easier to learn (Typological Prevalence Hypothesis - TPH). Prior work supports this hypothesis in phonology, morphology and syntax but has not addressed semantics. Using an Artificial Language Learning paradigm, we explore the learnability of semantic distinctions within the domain of evidentiality (i.e. the linguistic encoding of information sources). Our results support the TPH, since the most prevalent evidential system was learned best while the most rare evidentiality system yielded the worst learnability results. Furthermore, our results indicate that, cross- linguistically, indirect information sources seem to be marked preferentially (and acquired more easily) compared to direct sources. We explain this pattern in terms of the pragmatic need to mark indirect, potentially more unreliable sources over direct sources of information.

Languages partition semantic space into linguistic cate-gories in systematic ways. In this study, we investigatea semantic space which has received sustained attentionin theoretical linguistics: person. Person systems con-vey the roles entities play in the conversational context(i.e., speaker(s), addressee(s), other(s)). Like other lin-guistic category systems (e.g. color and kinship terms),not all ways of partitioning the person space are equallylikely. We use an artificial language learning paradigm totest whether typological frequency correlates with learn-ability of person paradigms. We focus on first personsystems (e.g., ‘I’ and ‘we’ in English), and test the predic-tions of a set of theories which posit a universal set of fea-tures (±exclusive, and ±minimal) to capture this space.Our results provide the first experimental evidence forfeature-based theories of person systems.

This thesis combines a number of methodologies and measures in order to address the processing of two types of negation (with not and un) in relation to each other and to non-negated affirmative meanings. The four investigations in this thesis target the processing of these meanings at different levels, namely the word level (Study 1) and sentence level (Studies 2–4). In two artificial language learning experiments, Study 1 focuses on the cognitive operations behind the processing of the negated meanings expressed by not and un in English. Negated and non-negated meanings are presented through artificial prefixes in order to do away with factors inherent to natural language (i.e., frequency and length). The meaning ranges associated with the negated types are manipulated and tested. The results reveal that the processing difficulty of these negation types can be modulated by the scopal ranges that they cover. This study can be taken to serve as a baseline for the future research on the processing of negated meanings that involve similar cognitive operations. Using event-related brain potentials (ERPs), Studies 2–4 target the time-course of processing negation in sentence comprehension. Affirmative, prefixally negated and sententially negated adjectives (authorized, unauthorized, not authorized) are presented in complex sentential contexts using both visual (Study 2 and Study 4) and auditory (Study 3) stimuli. The results of these investigations reveal a number of discursive factors that modulate the processing of these forms in sentence comprehension. More specifically, the results reveal that factors such as visual or auditory input, order of presentation, relevance and immediacy of the information presented all affect the processing of negated meanings in context. Overall, the findings of the studies in this thesis contribute new knowledge to a dynamic view of meaning and language processing and suggest that the processing of negated meanings, similar to many othe

Previous research on acquisition of noun class systems, such as grammatical gender, has shown that child learners rely dispro- portionately on phonological cues to class, even when compet- ing semantic cues are more reliable. Culbertson, Gagliardi, and Smith (2017) use artificial language learning experiments with adults to argue that over-reliance on phonology may be due to the fact that phonological cues are available first; learners base early representations on surface phonological dependen- cies, only later integrating semantic cues from noun meanings. Here, we show that child learners (6-7 year-olds) show this same sensitivity to early availability. However, we also find intriguing evidence of developmental changes in sensitivity to semantics; when both cues are simultaneously available chil- dren are more likely to rely on a phonology cue than adults. Our results suggest that early availability and a bias in favor of phonological cues may both contribute to children’s over- reliance on phonology in natural language acquisition.

Non-adjacent dependencies are ubiquitous in language, butdifficult to learn. Previous research has shown that the presenceof high variability between dependent items facilitateslearning. Yet what allows learning of non-adjacentdependencies even without high variability in interveningelements? One possibility is that learning non-adjacentdependencies highlights similar structures, allowing people tolearn new non-adjacent dependencies that are otherwisedifficult. In two studies, we show how being exposed tolearnable non-adjacent dependencies can change learners’sensitivity to novel non-adjacent regularities that are moredifficult to detect. These findings demonstrate a new way inwhich learning can build on and shape later learning aboutcomplex linguistic structure.

When asked to convey temporal concepts such as ‘yesterday’and ‘tomorrow’ via movements of a dot on a vertical bar,American undergraduates utilize analogical mappingsbetween spatial and temporal concepts. Previous work hasrevealed two different strategies, hypothesized to requirediffering amounts of artificial language exposure to learn.Different pairs of participants, when interacting about thesetime concepts, all settled on the same association betweenspatial magnitude and temporal duration, with largermovements used to convey temporal intervals of greaterduration. However, the association between particular spatiallocations and temporal concepts such as ‘past’ and ‘future’,elicited much more arbitrary solutions, where the mappingsdiffered across pairs of participants. These findings suggestedthat the duration mapping might be driven by mostly shared,initial cognitive biases, while contrasting mappings forpast/future result more clearly from extensive linguisticinteraction. Here we tested whether the brain respondsdifferently to duration mappings as compared to directionmappings by recording participants’ EEG as they learn amini-language that includes both kinds. ERPs time locked toEnglish words elicited larger amplitude N400 and P600 whenthey did not match the preceding signal than when they didmatch. The P600 results were larger and more robust for theduration than the direction stimuli, suggesting participantswere more sensitive to violations of the duration mappingscheme. These data support our hypothesis that people have acognitive bias for the duration mappings that supports theirearly emergence in the development of a semiotic system.

Learning syntax requires determining relations between the grammatical categories of words in the language, but learning those categories requires understanding the role of words in the syntax. In this study, we examined how this chicken and egg problem is resolved by learners of an artificial language comprising nouns, verbs, adjectives and case markers following syntactic rules. We found that the language could be acquired through cross-situational statistical correspondences with complex scenes and without explicit feedback, and that knowledge was maintained after 24 hours. Results also showed that verbs and word order were the first to be acquired, followed by nouns, adjectives and finally case markers. Interdependencies in learning were found for word order and verbs, and also for nouns, adjectives and case markers. Grammar and vocabulary can be acquired simultaneously, but with distinctive patterns of acquisition – grammar and the role of verbs first, then the vocabulary of other lexical categories.

Human languages contain very little unconditioned variation.In contexts where language learners are exposed to input thatcontains inconsistencies, they tend to regularise it, either byeliminating competing variants, or conditioning variant use onthe context. In the present study we compare regularisationbehaviour across linguistic levels, looking at how adult learn-ers respond to variability in morphology and word order. Ourresults suggest similar strengths in regularisation between lin-guistic levels given input languages whose complexity is com-parable.

Recent studies using artificial language learning have arguedthat the cross-linguistic frequency of harmonic word orderpatterns–in which heads are ordered consistently before or af-ter dependents across syntactic categories–reflects a cognitivebias (Culbertson, Smolensky, & Legendre, 2012; Culbertson& Newport, 2015a). These studies suggest that English speak-ing adults and children favor harmonic orders of nouns anddifferent nominal modifiers (adjectives, numerals). However,because they target English learners, whose native languageis harmonic in the nominal domain (Num-Adj-N), this pref-erence may be based on transfer rather than a universal biasfor harmony. We present new evidence from French-speakingchildren, whose native language is non-harmonic in this do-main (Num-N-Adj). Our results reveal clear effects of nativelanguage transfer, but also evidence that a harmonic pattern isfavored even in this population of learners.

Iconicity, i.e. resemblance between form and meaning, is awidespread feature of natural language vocabulary (Perniss,Thompson, & Vigliocco, 2010), and has been shown tofacilitate vocabulary acquisition (Imai, Kita, Nagumo, &Okada). But what kind of advantage does iconicity actuallygive? Here we use cross-situational learning (Yu & Smith,2007), to address the question for sound-shape iconicity (theso-called kiki-bouba effect, Ramachandran & Hubbard,2001). In contrast to Monaghan, Mattock, and Walker (2012),Experiment 1 suggests that the iconicity advantage comesfrom referential disambiguation rather than more efficientmemory encoding. Experiments 2 and 3 replicate this result,and moreover show that the kiki-bouba effect is roughlyequally strong for sharp and rounded shapes, a property thatclassic experiments were unable to confirm, and which hasimplication for the effect’s mechanism

We present the Retention and Recognition model (R&R), aprobabilistic exemplar model that accounts for segmentationin Artificial Language Learning experiments. We show thatR&R provides an excellent fit to human responses in threesegmentation experiments with adults (Frank et al., 2010),outperforming existing models. Additionally, we analyze theresults of the simulations and propose alternative explanationsfor the experimental findings.

Adjuncts and arguments exhibit different syntactic behaviours, but modelling this difference in minimalist syntax is challenging: on the one hand, adjuncts differ from arguments in that they are optional, transparent, and iterable, but on the other hand they are often strictly ordered, reflecting the kind of strict selection seen in argument application. The former properties mean the derivation proceeds the same way whether or not the adjuncts are present, but the latter means the derivation must know which adjuncts have already been adjoined, to avoid adjoining new ones out of order. This dissertation proposes a precise minimalist model of adjuncts that accounts for both behaviours. The second half considers the learnability of two closely related properties of adjuncts: their optionality and iterability. Many formal learning models predict a relationship between optionality and iterability, and any learning model of human language needs to be able to generalise from limited to indefinite repetition, since many languages include such sentences as "I really really really ... really love linguistics". All of the formal models I examine make this generalisation. A study of people learning an artificial language indicates that people also make this generalisation.

Many languages restrict their lexicons by OCP-Place, a phonotactic constraint against co-occurrences of consonants with shared [place] (e.g., McCarthy, 1986). While many previous studies have suggested that listeners have knowledge of OCP-Place and use this for speech processing, it is less clear whether they make reference to an abstract representation of this constraint. In Dutch, OCP-Place gradiently restricts non-adjacent consonant co-occurrences in the lexicon. Focusing on labial-vowel-labial co-occurrences, we found that there are, however, exceptions from the general effect of OCP-Labial: (A) co-occurrences of identical labials are systematically less restricted than co-occurrences of homorganic labials, and (B) some specific pairs (e.g., /pVp/, /bVv/) occur more often than expected. Setting out to study whether exceptions such as (A) and (B) had an effect on processing, the current study presents an artificial language learning experiment and a reanalysis of Boll-Avetisyan and Kager's (2014) speech segmentation data. Results indicate that Dutch listeners can use both knowledge of phonotactic detail and an abstract constraint OCP-Labial as a cue for speech segmentation. We suggest that whether detailed or abstract representations are drawn on depends on the complexity of processing demands.

Several studies examined cue competition in human learning by testing learners on a combination of conflicting cues rooting for different outcomes, with each cue perfectly predicting its outcome. A common result has been that learners faced with cue conflict choose the outcome associated with the rare cue (the Inverse Base Rate Effect, IBRE). Here, we investigate cue competition including IBRE with sentences containing cues to meanings in a visual world. We do not observe IBRE. Instead we find that position in the sentence strongly influences cue salience. Faced with conflict between an initial cue and a non-initial cue, learners choose the outcome associated with the initial cue, whether frequent or rare. However, a frequent configuration of non-initial cues that are not sufficiently salient on their own can overcome a competing salient initial cue rooting for a different meaning. This provides a possible explanation for certain recurring patterns in language change.

The role of processing constraints on sentence structure has been a topic of central interest in cognitive science. One proposal (Hawkins, 2004) suggests that language production system is organized to facilitate efficient parsing. We experimentally test this hypothesis using a miniature artificial language learning paradigm. Our findings support this account. Even though the input languages did not favor early placement of cues to grammatical function assignment (case and word order), participants used these cues in their own productions significantly more often in such a way as to allow early correct parsing commitments. This preference interacted with a bias to mark the less expected: Participants tended to use more case-marking in non-English OSV sentences. Our results underscore the potential of miniature artificial learning for language production research.

Current computational models of word learning make use of correspondences between words and observed referents, but as of yet cannot—as human learners do—leverage information regarding the meaning of other words in the lexicon. Here we develop a Bayesian framework for word learning that learns a lexicon from multiword utterances. In a set of three simulations we demonstrate this framework’s functionality, consistency with experimental work, and superior performance in certain learning tasks with respect to a Bayesian word leaning model that treats word learning as inferring the meaning of each word independently. This framework represents the first step in modeling the potential synergies between referential and distributional cues in word learning

It is argued that infants build a foundation for learning about the world through their incidental acquisition of the spatial and temporal regularities surrounding them. A challenge is that learning occurs across multiple contexts whose statistics can greatly differ. Two artificial language studies with 12-month-olds demonstrate that infants come prepared to parse statistics across contexts using the temporal and perceptual features that distinguish one context from another. These results suggest that infants can organize their statistical input with a wider range of features that typically considered. Possible attention, decision making, and memory mechanisms are discussed.

This study combined artificial language learning (ALL) with conventional experimental techniques to test whether statistical speech segmentation outputs are integrated into adult listeners' mental lexicon. Lexicalization was assessed through inhibitory effects of novel neighbors (created by the parsing process) on auditory lexical decisions to real words. Both immediately after familiarization and post-one week, ALL outputs were lexicalized only when the cues available during familiarization (transitional probabilities and wordlikeness) suggested the same parsing (Experiments 1 and 3). No lexicalization effect occurred with incongruent cues (Experiments 2 and 4). Yet, ALL differed from chance, suggesting a dissociation between item knowledge and lexicalization. Similarly contrasted results were found when frequency of occurrence of the stimuli was equated during familiarization (Experiments 3 and 4). Our findings thus indicate that ALL outputs may be lexicalized as far as the segmentation cues are congruent, and that this process cannot be accounted for by raw frequency., Journal Article, Research Support, Non-U.S. Gov't, SCOPUS: ar.j, info:eu-repo/semantics/published