Publications

Previous research has shown that content words are acquired earlier than function words, but the exact trajectory and order of function word acquisition has remained relatively understudied. In this study, we use the largest available child language corpora as well as Bayesian growth curve modeling to estimate the population level onset of production for more than 100 English function words. Our estimates suggest that for the large majority of function words, the earliest age of production lies between 12-24 months. A linear regression found longer function words as measured by the number of phonemes and function words with higher Mean Length of Utterance (MLU) have later estimated onset age of production. We did not find a significant effect of frequency in child directed-speech on onset age of production for function words. Overall these results point to early emergence of abstract functional morphemes with production limitations as the main bottleneck. 

Building an NLP training set for understudied languages, including Indigenous and endangered languages, often faces challenges due to varying degrees of resource limitations in the speaker communities. What are some reasonable approaches for training set construction in these cases? We address this question with POS tagging as the test case. Although many might consider POS tagging “a solved problem”, it remains a crucial task for descriptive linguistics and language documentation and requires laborious manual annotation. Drawing data from 12 language families, we compare in-context learning, active learning (AL), and random sampling. Our results suggest: (1) for communities whose language data can be ethically shared with an API, using only 1,000 randomly sampled tokens as prompt examples, the proprietary GPT-4.1-mini can deliver desirable performance (F1>0.83) on par with that from a training set of thousands of tokens in AL iterations; (2) in cases where communities prefer not to share data, 4,500-5,500 tokens selected from AL can yield reasonable results at a pace statistically significantly faster than random sampling, evidenced by growth curve modeling.

Humans are capable of language learning with or without instruction. While “instruction” is typically associated with formal education, it can be more broadly defined as any intervention in the learning process with the intention of improving the process or the outcome. This intervention may be in the selection of the linguistic data that informs language learning, it may be the design of language learning tasks, the provision of feedback that guides the learner, or the presentation of meta-linguistic generalizations that consciously inform the learner about the grammar of the language. It can be “indirect” by simply providing the environment that facilitates learning, or it can be “direct” by providing pre-defined learningoutcomes, a structured syllabus, or metalinguistic descriptive generalizations (Ellis et al. 2009). A fundamental question in second language acquisition is whether instruction improves language learning; and if so, what type of instruction is optimal for language learning (Loewen 2020)?

How does linguistic negation develop in early child language? Prior research has suggested that abstract and context-general negation develops from concrete and context-specific communicative functions such as rejection, prohibition, or nonexistence in fixed and ordered stages. The evidence for the emergence of these functions in stages is mixed, however, leaving the possibility that negation starts as an abstract concept that can serve multiple specific functions from the beginning, and that the development of the different functions starts more or less simultaneously depending on the early communicative environment. Leveraging automatic annotations of large-scale child speech corpora in English and growth-curve modeling, we examine children’s production of seven negative constructions that tend to convey communicative functions previously discussed in the literature. We also investigate children’s discourse-level negative responses (saying no) to parents’ utterances with the same constructions as a proxy for children’s comprehension. We do not find strong evidence for population-level stages in children’s development of negation. Instead, the results of our growth-curve modeling suggest that for our measures of comprehension and production, children’s ability to negate different constructions likely emerges around 18–22 months of age. Our results complement and confirm recent findings in experimental studies on children’s comprehension of negation.

Disjunction in natural language alternates between an inclusive reading (A or B or Both) and an exclusive  reading (A or B but not Both). Traditional accounts of this ambiguity focus on scalar implicature as the source of disjunction exclusivity, a process whereby Gricean reasoning over Horn scales strengthens the baseline inclu-sive  reading  to  an  implied  exclusive  reading  (Grice,  1978;  Horn,  1972;  Gazdar, 1980). Despite nearly all theories acknowledging that other factors likely play a role in  the  generation  of  exclusivity  implications,  non-implicature  factors  have  received comparatively  little  attention. Across  four  experiments  we  tested  two  such  non  im-plicature factors, prior compatibility and syntactic category, finding that both play a role in speaker interpretations of disjunctive sentences. Additionally, by drawing our stimuli in the first two experiments from the prior literature, we found evidence that previous  research  on  disjunction,  while  accurately  identifying  the  key  role  of  scalar implicatures, may be overestimating the effect size thereof due to a failure to control for non-implicature factors.

How can modern neural networks like language models be useful to the field of language acquisition, and more broadly cognitive science, if they are not a priori designed to be cognitive models? As developments towards natural language understanding and generation have improved leaps and bounds, with models like GPT-4, the question of how they can inform our understanding of human language acquisition has re-emerged. As such, it is critical to examine how in practice linking hypotheses between models and human learners can be safely established. To address these questions, we propose a model taxonomy, including four modelling approaches, each having differing goals, from exploratory hypothesis generation to hypothesis differentiation and testing. We show how the goals of these approaches align with the overarching goals of science and linguistics by connecting our taxonomy to the realist versus instrumentalist approaches in philosophy of science. We survey recent work having adopted each of our modelling approaches and address the importance of computational modelling in language acquisition studies.

In English binomial expressions, ‚"bread and butter"‚ is preferred over ‚"butter and bread". Morgan & Levy (2015) show that for these types of expressions, frequently used phrases tend to have stronger, more extreme preferences. In contrast, there is roughly an equal preference for ‚"bishops and rooks‚" versus ‚"rooks and bishops‚" a much less common pairing. This paper extends this research to the concept of Mandarin elastic word length, a phenomenon in which most Mandarin words have long and short forms.  We find evidence for frequency-dependent regularization in the elastic length of Noun-Noun compounds in Chinese, demonstrating that frequency-dependent regularization extends to structures with more than two alternations and to languages other than English.

How does abstract linguistic negation develop in early child language? Previous research has suggested that abstract negation develops in stages and from more concrete communicative functions such as rejection, prohibition, or non-existence. The evidence for the emergence of these functions in stages is mixed, however, leaving the possibility that negation is an abstract concept from the beginning that can serve multiple specific functions depending on early communicative environment. Leveraging automatic annotations of large-scale child speech corpora in English, we examine the production trajectories of seven negative constructions that tend to convey communicative functions previously discussed in the literature. The results demonstrate the emergence and gradual increase of these constructions in child speech within the age range of 18-36 months. Production mostly remains stable, regular, and close to parents’ levels after this age range. These findings are consistent with two hypotheses: first, that negation starts as an abstract concept that can serve multiple functions from the beginning; and second, that negation develops in stages from specific communicative functions but this development is early and quick, leaving our corpus methods incapable of detecting them from the available corpus data.