Artificial Language Learning Experiments Research Articles

One Cue's Loss Is Another Cue's Gain-Learning Morphophonology Through Unlearning.

A word often expresses many different morphological functions. Which part of a word contributes to which part of the overall meaning is not always clear, which raises the question as to how such functions are learned. While linguistic studies tacitly assume the co-occurrence of cues and outcomes to suffice in learning these functions (Baer-Henney, Kügler, & van de Vijver, 2015; Baer-Henney & van de Vijver, 2012), error-driven learning suggests that contingency rather than contiguity is crucial (Nixon, 2020; Ramscar, Yarlett, Dye, Denny, & Thorpe, 2010). In error-driven learning, cues gain association strength if they predict a certain outcome, and they lose strength if the outcome is absent. This reduction of association strength is called unlearning. So far, it is unclear if such unlearning has consequences for cue-outcome associations beyond the ones that get reduced. To test for such consequences of unlearning, we taught participants morphophonological patterns in an artificial language learning experiment. In one block, the cues to two morphological outcomes-plural and diminutive-co-occurred within the same word forms. In another block, a single cue to only one of these two outcomes was presented in a different set of word forms. We wanted to find out, if participants unlearn this cue's association with the outcome that is not predicted by the cue alone, and if this allows the absent cue to be associated with the absent outcome. Our results show that if unlearning was possible, participants learned that the absent cue predicts the absent outcome better than if no unlearning was possible. This effect was stronger if the unlearned cue was more salient. This shows that unlearning takes place even if no alternative cues to an absent outcome are provided, which highlights that learners take both positive and negative evidence into account-as predicted by domain general error-drivenlearning.

Predictability and Variation in Language Are Differentially Affected by Learning and Production.

General principles of human cognition can help to explain why languages are more likely to have certain characteristics than others: structures that are difficult to process or produce will tend to be lost over time. One aspect of cognition that is implicated in language use is working memory-the component of short-term memory used for temporary storage and manipulation of information. In this study, we consider the relationship between working memory and regularization of linguistic variation. Regularization is a well-documented process whereby languages become less variable (on some dimension) over time. This process has been argued to be driven by the behavior of individual language users, but the specific mechanism is not agreed upon. Here, we use an artificial language learning experiment to investigate whether limitations in working memory during either language learning or language production drive regularization behavior. We find that taxing working memory during production results in the loss of all types of variation, but the process by which random variation becomes more predictable is better explained by learning biases. A computational model offers a potential explanation for the production effect using a simple self-primingmechanism.

Naturalness is gradient in morphological paradigms: Evidence from positional splits

Agreement markers that refer to the same feature or argument tend to be found in the same position (e.g., all subject agreement markers as suffixes, all object agreement markers as prefixes). However, little is known about the exceptions to this trend: cases where different values of the same feature are marked in different positions in the word (i.e., positional splits). In this study, we explore the positional properties of subject and object person-number agreement markers in a phylogenetically diverse sample of 227 languages. We find that the recurrence of a positional split is proportional to its degree of naturalness, that is, to the amount of shared feature values amongst the forms with the same positional arrangement. Natural patterns (e.g., where prefixal forms all share SG and suffixal forms all share PL) are over-represented in natural languages compared to a random baseline. The most unnatural patterns are underrepresented, and splits with an intermediate level of unnaturalness occur at around chance levels. We hypothesise that this graded bias for naturalness is grounded in a preference for morphological similarity amongst semantically similar forms during language learning. To test this hypothesis we conducted two online artificial language learning experiments where we trained and tested participants on person-number verbal agreement paradigms of different sizes with positional splits of different degrees of naturalness. We found that their relative learnability is also gradient, again proportional to the amount of feature value overlap, thus matching the observed cross-linguistic tendencies. Our findings support the notion that semantic similarity shapes the evolution of morphological structure in person-number verbal agreement systems and that it does so in a gradient way.

Modeling harmony biases in learning exceptions to vowel harmony

Artificial language learning experiments typically show non-categorical results after training on categorical data. This is generally due to incomplete learning, but these results can also reveal biases. One example is that participants trained on a vowel harmony language with alternating and non-alternating affixes prefer the non-alternating affix in harmonic contexts (Finley 2021). In this paper, I show that (i) the preference for harmonic items in non-alternating affixes replicates for remote (online) data collection, and (ii) that this effect can be modeled with MaxEnt Harmonic Grammar. In Harmonic Grammar, the harmony score of each candidate determines its grammaticality, and the probability of surfacing. Because non-alternating affixes that satisfy vowel harmony have higher harmony scores than non-alternating affixes that violate harmony, harmonic candidates will be more likely to surface than disharmonic candidates, even when both types of items surface at levels greater than expected by chance. The theoretical and methodological implications for these results are discussed

Acquiring recursive structures through distributional learning

ABSTRACT Languages differ regarding the depth, structure, and syntactic domains of recursive structures. Even within a single language, some structures allow infinite self-embedding while others are more restricted. For example, when expressing ownership relation, English allows infinite embedding of the prenominal genitive -s, whereas the postnominal genitive of is much more restricted. How do speakers learn which specific structures allow infinite embedding and which do not? The distributional learning proposal suggests that the recursion of a structure (e.g., X1’s-X2 ) is licensed if the X1 position and the X2 position are productively substitutable in non-recursive input. The present study tests this proposal with an artificial language learning experiment. We exposed adult participants to X1-ka-X2 strings. In the productive condition, almost all words attested in X1 position were also attested in X2 position; in the unproductive condition, only some were. We found that, as predicted, participants from the productive condition were more likely to accept unattested strings at both one- and two-embedding levels than participants from the unproductive condition. Our results suggest that speakers can use distributional information at one-embedding level to learn whether or not a structure is recursive.

“Natural” stress patterns and dependencies between edge alignment and quantity sensitivity

We conducted an artificial language learning experiment to study learning asymmetries that might reveal latent preferences relating to, and any dependencies between, the edge alignment and quantity sensitivity (QS) parameters in stress patterning. We used a poverty of the stimulus approach to teach American English speakers an unbounded QS stress rule (stress a single CV: syllable) and either a left- or right-aligning QI rule if only light syllables were present. Forms with two CV: syllables were withheld in the learning phase and added in the test phase, forcing participants to choose between left- and right-aligning options for the QS rule. Participants learned the left- and right-edge QI rules equally well, and also the basic QS rule. Response patterns for words with two CV: syllables suggest biases favoring a left-aligning QS rule with a left-edge QI default. Our results also suggest that a left-aligning QS pattern with a rightedge QI default was least favored. We argue that stress patterns shown to be preferred based on evidence from ease-of-learning and participants’ untrained generalizations can be considered more natural than less favored opposing patterns. We suggest that cognitive biases revealed by artificial stress learning studies may have contributed to shaping stress typology.

More or Less Unnatural: Semantic Similarity Shapes the Learnability and Cross-Linguistic Distribution of Unnatural Syncretism in Morphological Paradigms.

Morphological systems often reuse the same forms in different functions, creating what is known as syncretism. While syncretism varies greatly, certain cross-linguistic tendencies are apparent. Patterns where all syncretic forms share a morphological feature value (e.g., first person, or plural number) are most common cross-linguistically, and this preference is mirrored in results from learning experiments. While this suggests a general bias towards natural (featurally homogeneous) over unnatural (featurally heterogeneous) patterns, little is yet known about gradients in learnability and distributions of different kinds of unnatural patterns. In this paper we assess apparent cross-linguistic asymmetries between different types of unnatural patterns in person-number verbal agreement paradigms and test their learnability in an artificial language learning experiment. We find that the cross-linguistic recurrence of unnatural patterns of syncretism in person-number paradigms is proportional to the amount of shared feature values (i.e., semantic similarity) amongst the syncretic forms. Our experimental results further suggest that the learnability of syncretic patterns also mirrors the paradigm's degree of feature-value similarity. We propose that this gradient in learnability reflects a general bias towards similarity-based structure in morphological learning, which previous literature has shown to play a crucial role in word learning as well as in category and concept learning more generally. Rather than a dichotomous natural/unnatural distinction, our results thus support a more nuanced view of (un)naturalness in morphological paradigms and suggest that a preference for similarity-based structure during language learning might shape the worldwide transmission and typological distribution of patterns of syncretism.

Syntax Matters: Exploring the Effect of Linguistic Similarity in Third Language Acquisition

AbstractOver the last two decades, the question of to which linguistic cues learners pay attention when they decode a new language has been subject to controversy in the field of third language (L3) acquisition. In this article, we present an artificial language learning experiment that investigated how lexical and syntactic similarities between an artificial L3 and preexisting grammars impact crosslinguistic influence at the very beginning of the acquisition process. We exposed four groups of 30 Norwegian–English bilinguals each to one of four L3s and gave them training in that L3. The participants gave forced‐choice acceptability judgments on pairs of nonsubject‐initial declarative clauses that differed in word order, one grammatical in English, the other grammatical in Norwegian. The participants had not been exposed to nonsubject‐initial declaratives during the exposure and training phases to avoid confounds with learning. The results showed that both lexical and syntactic similarities affect crosslinguistic influence. We discuss this result considering contemporary accounts of L3 acquisition.

Generalization to Novel Consonants: Place Versus Voice.

In traditional, generative phonology, sound patterns are represented in terms of abstract features, typically based on the articulatory properties of the sounds. The present study makes use of an artificial language learning experiment to explore when and how learners extend a novel phonological pattern to novel segments. Adult, English-speaking learners were exposed to a spirantization pattern in which a stop became a fricative between two vowels (e.g., /bib/ + /o/ ➔ [bivo]). Participants were trained on spirantization for two of four possible stop-fricative pairs, and were tested on their generalization to the held-out segments. Two groups of participants were trained on items based on voicing (e.g., the Voiced condition was trained on /b/ ➔ [v], and /d/ ➔ [z], and tested on /p/ ➔ [f], and /t/ ➔ [s]), and two groups of participants were trained on items based on place of articulation (e.g., the Labial condition was trained on /b/ ➔ [v], and /p/ ➔ [f] and tested on /t/ ➔ [s], and /d/ ➔ [z]). Participants in both Place and Voice conditions were successful at learning and generalizing the spirantization pattern to novel segments, but rates of generalization were higher in the Voice conditions. These results support a similarity-based approach to generalization, particularly one that takes into account articulatorily-based features and natural classes. Implications for phonological theory are discussed.

The effect of lengthening aspiration on speech segmentation.

Listeners tend to perceive longer vowels as word-final and include the aspiration of a syllable-onset consonant in metalinguistically judging the duration of the following vowel. It may be hypothesized that longer aspiration is interpreted as extra vowel duration and hence, a word-finality cue. Alternatively, if aspiration is perceived as part of the onset, longer aspiration should be interpreted as word-initial, consistent with previous findings on consonant lengthening [White, Mattys, Stefansdottir, and Jones (2015). J. Acoust. Soc. Am. 138(2), 1214-1220]. An artificial language learning experiment with Taiwanese Southern Min listeners showed that lengthening aspiration in word-initial but not word-final syllables improved speech segmentation, supporting the second but not the first hypothesis.

The Impact of Information Structure on the Emergence of Differential Object Marking: An Experimental Study.

Many languages exhibit differential object marking (DOM), where only certain types of grammatical objects are marked with morphological cases. Traditionally, it has been claimed that DOM arises as a way to prevent ambiguity by marking objects that might otherwise be mistaken for subjects (e.g., animate objects). While some recent experimental work supports this account, research on language typology suggests at least one alternative hypothesis. In particular, DOM may instead arise as a way of marking objects that are atypical from the point of view of information structure. According to this account, rather than being marked to avoid ambiguity, objects are marked when they are given (already familiar in the discourse) rather than new. Here, we experimentally investigate this hypothesis using two artificial language learning experiments. We find that information structure impacts participants’ object marking, but in an indirect way: atypical information structure leads to a change in word order, which then triggers increased object marking. Interestingly, this staged process of change is compatible with documented cases of DOM emergence. We argue that this process is driven by two cognitive tendencies. First, a tendency to place discourse given information before new information, and second, a tendency to mark noncanonical word order. Taken together, our findings provide corroborating evidence for the role of information structure in the emergence of DOM systems.

Nobody Doesn\u2019t Like Negative Concord

Languages vary with respect to whether sentences with two negative elements give rise to double negation or negative concord meanings. We explore an influential hypothesis about what governs this variation: namely, that whether a language exhibits double negation or negative concord is partly determined by the phonological and syntactic nature of its negative marker (Zeijlstra 2004; Jespersen 1917). For example, one version of this hypothesis argues that languages with affixal negation must be negative concord (Zeijlstra 2008). We use an artificial language learning experiment to investigate whether English speakers are sensitive to the status of the negative marker when learning double negation and negative concord languages. Our findings fail to provide evidence supporting this hypothesised connection. Instead, our results suggest that learners find it easier to learn negative concord languages compared to double negation languages independently of whether the negative marker is an adverb or an affix. This is in line with evidence from natural language acquisition (Thornton et al. 2016).

Modeling word and morpheme order in natural language as an efficient trade-off of memory and surprisal.

Memory limitations are known to constrain language comprehension and production, and have been argued to account for crosslinguistic word order regularities. However, a systematic assessment of the role of memory limitations in language structure has proven elusive, in part because it is hard to extract precise large-scale quantitative generalizations about language from existing mechanistic models of memory use in sentence processing. We provide an architecture-independent information-theoretic formalization of memory limitations which enables a simple calculation of the memory efficiency of languages. Our notion of memory efficiency is based on the idea of a memory-surprisal trade-off: A certain level of average surprisal per word can only be achieved at the cost of storing some amount of information about the past context. Based on this notion of memory usage, we advance the Efficient Trade-off Hypothesis: The order of elements in natural language is under pressure to enable favorable memory-surprisal trade-offs. We derive that languages enable more efficient trade-offs when they exhibit information locality: When predictive information about an element is concentrated in its recent past. We provide empirical evidence from three test domains in support of the Efficient Trade-off Hypothesis: A reanalysis of a miniature artificial language learning experiment, a large-scale study of word order in corpora of 54 languages, and an analysis of morpheme order in two agglutinative languages. These results suggest that principles of order in natural language can be explained via highly generic cognitively motivated principles and lend support to efficiency-based models of the structure of human language. (PsycInfo Database Record (c) 2021 APA, all rights reserved).

Production and Substantive Bias in Phonological Learning

The role of inductive biases has been actively examined in work on phonological learning. While previous studies systematically supported a structural bias hypothesis, i.e., patterns with simpler phonological featural descriptions are easier to learn, the results have been mixed for a substantive bias hypothesis, i.e., phonetically motivated patterns are easier to learn. This study explores an explanation for the uncertain status of substantive bias in phonological learning. Among the aspects of phonetic substance, we focus on articulatory factors. We hypothesize that practice producing phonological patterns makes salient to learners the articulatory factors underlying articulatorily (un-)grounded patterns. An artificial language learning experiment was conducted to test the learning of postnasal (de)voicing, a pattern which is primarily grounded on articulatory components. We examine the role of production in the learning of articulatorily grounded (postnasal voicing) vs. ungrounded patterns (postnasal devoicing), by comparing the outcomes of perception-only vs. perception-with-production learning contexts, both in categorical and variable pattern learning conditions. The results show evidence for a production effect, but it was restricted to certain contexts, namely those involving a higher level of uncertainty and for languages exhibiting dominant natural patterns. We discuss the implications of our findings for phonological learning and language change.

Familiarity, consistency, and systematizing in morphology

Language learning involves exposure to inconsistent systems - that is, systems where multiple patterns or methods exist to mark some meaning. Inconsistent systems often change to be more regular over time - they become systematized. However, some recent studies have reported that learners tend to reproduce inconsistency in the input, leading to models in which the language learning mechanism is basically preservatory. We ran an artificial language learning experiment using a novel paradigm to extend our understanding of systematizing versus preservatory mechanisms in language learning. Participants were taught two number marking systems, either completely consistently (the probability P of the system is 1.00) or inconsistently (with P = 0.875 for one system and P = 0.125 for the other, and so on for P = 0.75 and P = 0.625). One marking system was a plural-marking system. The other was a typologically rare singulative-marking system. When generalizing to novel items, participants produced more regular output patterns overall for more consistent input conditions than for less consistent ones, and more for the plural-marking conditions than for the singulative-marking conditions. For the singulative-marking conditions, the inter-participant variation was much greater than for the plural-marking ones; some individuals systematized towards the more familiar pattern, some systematized towards the less familiar pattern and some were not significantly different from probability-matching. We analyze the variation in relation to current statistical learning models, showing that preservatory learning models, as well as all models with a single free parameter, fail to capture our results. We show how a model with two free parameters in which individuals can vary in their propensity to systematize in any given situation is more successful. We also discuss implications for the theory of language change.

Coarticulation and learnability of transparent vowels in vowel harmony

The representations of transparent vowels in vowel harmony have been of interest to phonologists because of the challenges they pose for constraints on locality and complexity. One proposal is that transparent vowels in back vowel harmony may be intermediate between front and back. The present study uses two artificial language learning experiments to explore the psychological reality of acoustic differences in transparent vowels in back vs. front vowel contexts. Participants were exposed to a back/round vowel harmony language with a neutral vowel that was spliced so that the F2 was lower in back vowel contexts and higher in front vowel contexts (the Natural condition) or the reverse (the Unnatural condition). While only participants in the Natural condition of Experiment 1 were able to learn the behavior of the transparent vowel relative to a No-Training control, there was no difference between the Natural and Unnatural conditions. In Experiment 2, only participants in the Natural condition learned the vowel harmony pattern, though there were no significant differences between the two conditions. No condition successfully learned the behavior of the transparent vowel in Experiment 2. These results suggest that the effects of small differences in the F2 value of transparent back vowels on learnability are minimal.

Language learning as uncertainty reduction: The role of prediction error in linguistic generalization and item-learning

Discriminative theories frame language learning as a process of reducing uncertainty about the meaning of an utterance by discriminating informative from uninformative cues via the mechanisms of prediction error and cue competition. Previous work showed that discriminative learning is affected by the order in which information is presented during language learning. Specifically, learning suffixes, where complex stems precede affixes, promotes better generalization than prefixing, which tends to promote better item-learning instead. We explored this in two large-scale web-based artificial language learning experiments with adult learners (total N = 434), as well as two computational simulations implementing a discriminative learning model. While we did not find an overall benefit of suffixing over prefixing in generalization, consistent with our theoretical and computational predictions, we found that participants in the prefix condition were unable to discriminate between frequent, but uninformative cues and low-frequency, informative cues. This resulted in them being more likely to show incorrect overgeneralization of that feature for low frequency test items than participants in the suffix condition. We did not find a benefit of prefixing in item learning (although there was overall better item-learning of low type-frequency items), which we discuss in terms of the methodological limitations of our empirical paradigm. Taken together, these results underline the crucial role prediction error plays in learning linguistic generalization, and have implications for how generalization interacts with item-learning.

The Influence of Different Prosodic Cues on Word Segmentation.

A prerequisite for spoken language learning is segmenting continuous speech into words. Amongst many possible cues to identify word boundaries, listeners can use both transitional probabilities between syllables and various prosodic cues. However, the relative importance of these cues remains unclear, and previous experiments have not directly compared the effects of contrasting multiple prosodic cues. We used artificial language learning experiments, where native German speaking participants extracted meaningless trisyllabic “words” from a continuous speech stream, to evaluate these factors. We compared a baseline condition (statistical cues only) to five test conditions, in which word-final syllables were either (a) followed by a pause, (b) lengthened, (c) shortened, (d) changed to a lower pitch, or (e) changed to a higher pitch. To evaluate robustness and generality we used three tasks varying in difficulty. Overall, pauses and final lengthening were perceived as converging with the statistical cues and facilitated speech segmentation, with pauses helping most. Final-syllable shortening hindered baseline speech segmentation, indicating that when cues conflict, prosodic cues can override statistical cues. Surprisingly, pitch cues had little effect, suggesting that duration may be more relevant for speech segmentation than pitch in our study context. We discuss our findings with regard to the contribution to speech segmentation of language-universal boundary cues vs. language-specific stress patterns.

The differential effects of vowel and onset consonant lengthening on speech segmentation: Evidence from Taiwanese Southern Min.

A review of previous speech segmentation research suggests the prediction that listeners of Taiwanese Southern Min (TSM), a lexical tone language, would exploit vowel lengthening and syllable-onset consonant lengthening to locate word ends and beginnings, respectively. Yet, correlations between segment duration and tone identity in tone languages along with some TSM-specific phonological phenomena may work against such use. Two artificial language learning experiments examined TSM listeners' use of the lengthening cues. The listeners heard the words of an artificial language (e.g., /ba.nu.me/) repeated continuously and identified them in a subsequent two-alternative forced-choice test. Experiment I revealed that their segmentation benefits from and only from word-initial onset lengthening or word-final vowel lengthening, supporting the prediction. Experiment II further demonstrated that these two cues in combination synergistically support segmentation at least when compared to word-initial onset lengthening alone, consistent with previous findings regarding complementary cues. These results furnish additional evidence that vowel and onset consonant lengthening affect segmentation in different ways, possibly reflecting a functional division between vowels and consonants that is supported by some prosody-computing mechanism. Additionally, vowel lengthening seems to affect segmentation to a greater extent than onset consonant lengthening. Possible explanations for this and further issues are discussed.

Is Regularization Uniform across Linguistic Levels? Comparing Learning and Production of Unconditioned Probabilistic Variation in Morphology and Word Order

ABSTRACT Languages exhibit variation at all linguistic levels, from phonology, to the lexicon, to syntax. Importantly, that variation tends to be (at least partially) conditioned on some aspect of the social or linguistic context. When variation is unconditioned, language learners regularize it – removing some or all variants, or conditioning variant use on context. Previous studies using artificial language learning experiments have documented regularizing behavior in the learning of lexical, morphological, and syntactic variation. These studies implicitly assume that regularization reflects uniform mechanisms and processes across linguistic levels. However, studies on natural language learning and pidgin/creole formation suggest that morphological and syntactic variation may be treated differently. In particular, there is evidence that morphological variation may be more susceptible to regularization. Here we provide the first systematic comparison of the strength of regularization across these two linguistic levels. In line with previous studies, we find that the presence of a favored variant can induce different degrees of regularization. However, when input languages are carefully matched – with comparable initial variability, and no variant-specific biases – regularization can be comparable across morphology and word order. This is the case regardless of whether the task is explicitly communicative. Overall, our findings suggest an overarching regularizing mechanism at work, with apparent differences among levels likely due to differences in inherent complexity or variant-specific biases. Differences between production and encoding in our tasks further suggest this overarching mechanism is driven by production.