United States

Understanding how children process ambiguous words is a challenge because sense disambiguation depends on sentence context bottom-up and top-down aspects. Here, we seek insight into this phenomenon by investigating how such a competence might arise in large distributional learners (Transformers) that purport to acquire sense representations from language input in a largely unsupervised fashion. We investigated how sense disambiguation might be achieved using model representations derived from naturalistic child-directed speech. We tested a large pool of Transformer models, varying in their pretraining input size/nature as well as the size of their parameter space. Tested across three behavioral experiments from the developmental literature, we found that these models capture some essential properties of child sense disambiguation, although most still struggle in the more challenging tasks with contrastive cues. We discuss implications for both theories of word learning and for using Transformers to capture child language processing.

CogSci 2023

Modeling Children&#39;s Word Sense Disambiguation with Large Language Models: Insights and Challenges

ECAI2020 featured a full programme of technical papers, as well as many AI-related events of interest for researchers, students and all attendants who are interested on contemporary AI, including:




**An extensive programme of workshops, tutorials, and invited speakers**

**The Prestigious Applications of Intelligent Systems conference (PAIS)**

**The Frontiers in AI track sessions**

**The Lab To Market event**

**The Women in AI time**

**The EU Challenges forum**

… and many more!

ECAI 2020

ECAI 2020 featured a full programme of technical papers, as well as many AI-related events of interest for researchers, students and all attendants who are interested on contemporary AI.

Modeling Children's Word Sense Disambiguation with Large Language Models: Insights and Challenges

poster

The 45th Annual Meeting of the Cognitive Science Society will take place in Sydney, Australia as a modified hybrid conference. Presentations from the main-session room will be live-streamed every day for all events. This will include all invited talks and award sessions. Recordings of all CogSci 2023 sessions will be made available for on-demand viewing after the conference.<br>
The in-person component will be held at the International Convention Centre Sydney at Darling Harbour. The Conference will take place Wednesday July 26th – Saturday July 29th, 2023 (Wednesday, July 26th is the Workshops Day).

Join us!

<br>[**The Program at a Glance can be found here**](https://cognitivesciencesociety.org/wp-content/uploads/2023/07/COgSci-PAAG-V2.pdf)
<br>[**The Detailed Program can be found here**](https://online.fliphtml5.com/ebtyf/mswz/)
<br><br>

#### **CogSci 2023 Program Book**
<iframe style="width:700px;height:400px" src="https://online.fliphtml5.com/ebtyf/tzgk/"  seamless="seamless" scrolling="no" frameborder="0" allowtransparency="true" allowfullscreen="true" ></iframe>
<br>

The 45th Annual Meeting of the Cognitive Science Society presents the latest research across cognitive science and highlights the theme of Cognition in Context.

Artificial grammar learning (AGL) experiments are commonly used to assess implicit learning. However, in these tasks participants may acquire explicit rather than implicit knowledge. We combined an AGL task with measures of metacognition to assess whether participants acquired explicit knowledge of the grammar. Participants were initially exposed to sequences of visual symbols generated by an artificial grammar, followed by a testing phase in which they classified sequences as either grammatical or ungrammatical. On a subset of trials (choice trials) participants were given the option to either take or skip this classification. If participants had explicit knowledge of the grammar, they should be more likely to take trials when they are confident and skip when they are not. Participants learned the grammar, and performed better in choice trials than in forced trials, demonstrating explicit knowledge of the grammar. These results suggest that AGL paradigms might elicit explicit rather than implicit learning.

Using measures of metacognition to assess the acquisition of implicit and explicit knowledge in artificial grammar learning tasks

Morphosyntactic rules are central to compositional structure in language, and therefore crucial in language acquisition. Previous research on L1 and L2 acquisition suggests that active production leads learners to notice linguistic generalisations more than passive learning does. Here, we explore whether production helps learners to access less obvious generalisations when more than one analysis is possible. We taught participants an artificial language compatible with two grammatical analyses: one that marks semantic roles using word order, the other using case suffixes. After training, half of the participants do a comprehension task, and the other half do a production task. We predict that the comprehension group will adopt the English-like word order analysis, while the production group will be more likely to learn the recurring case suffixes and thus adopt the alternative analysis. Overall, our results shed light on the role of retrieval and production for the acquisition of morphosyntactic generalisations.

Does production facilitate learning morphosyntactic generalisations?

Previous studies suggest that speech categorization is established during the first year of life (Werker & Tees, 1984). However, recent work suggests protracted development through adolescence (McMurray et al., 2018). While the early development has been thought to be driven by the statistical properties of the input, it is unclear if this late phase is also input driven. We asked whether being exposed to diverse linguistic input leads to different patterns of speech categorization. Visual Analogue Scaling task and social network questionnaires were used to measure speech categorization and the effects of linguistic diversity in children’s and adults’ social environments. Results suggest that the development of speech perception is not finished during the first year of life. It continues to be shaped by the developmental changes across the lifespan (McMurray et al., 2018). Importantly, diverse linguistic experience makes both children and adults more flexible in their speech categorization.

Network Diversity Leads to More Flexibility in Speech Categorization

Language learning emerges from dynamic communicative exchanges characterized by two key dimensions. First, children are not simply passive observers, but instead actively seek new information. Second, caregivers scaffold the learning environment. We will present results from a preregistered study (target N=90; 36-72 mos; preregistration: https://osf.io/s24eb/?view_only=7cbed0ef12354103b5fe7a897d13f694; current sample size: N=51) investigating how caregiver scaffolding and children’s information-seeking mutually contribute to word learning. We introduce a novel word learning paradigm that independently varies whether children and/or their caregivers can structure input in a joint Learning Phase and measure children’s word learning in a pre-post test design. We predict that (1) children and caregivers will actively structure learning events and (2) children will benefit when they themselves and their caregivers can structure the Learning Phase. Our results will provide insight into how children’s information-seeking strategies interact with caregivers’ scaffolding to create informative word learning moments.

The dynamics of child-driven active information-seeking and caregiver scaffolding during word learning

The early moments of compound and pseudocompound visual word recognition were investigated by probing their “constituent” concepts (e.g., BED in bedroom and CAP in capsize). This was achieved by concomitantly presenting target words in one visual field (left or right, projected to the right or left hemisphere, respectively) and a picture representing the referent of either the first or second “constituent” in the opposing visual field. The stimuli were presented for 133 ms followed by a backward mask and participants judged whether the word and picture were related to each other. The experimental manipulations consisted of target word type (compound or pseudocompound), word complexity (whole word or “constituent”), probed constituent position (first or second “constituent”), and word projection (left or right hemisphere). Our results suggest that the “constituents” of compounds and pseudocompounds are conceptually accessed. We discuss the implications of our findings for the nature of the visual word recognition system.

Semantic access to constituents of compounds and pseudocompounds: Evidence from dichoptic presentation

In this study we investigate semantic retrieval strategies in order to inform discussions about cognitive mechanisms underlying divergent thinking and creativity. Relying on a verbal fluency task, we map the particular associative strategies participants engage and how they predict their performance. The study starts from the assumption that during divergent thinking processes, participants move along a path through a semantic space, and that each step is prompted by an associative strategy taking them from one word to the next. There are, however, a number of such strategies, and we predict that the outcome of the process is contingent on participants’ engagement of and shift between strategies. The study consists of a two-part elicitation paradigm where participants first conduct a verbal fluency task and then are guided through a meta-cognitive retrieval process to individuate the strategies employed in the task. We report significant correlations between the engagement of associative strategies and outcome measures of divergent thinking in terms of originality, flexibility and fluency.

Semantic Retrieval Strategies in Divergent Thinking

Young children’s intuitions about the meanings of novel nouns reveal foundational biases in language learning, such as biases to generalize nouns to objects of the same shapes, to whole—as opposed to parts of—objects, and to objects from the same taxonomic category. Nevertheless, existing work on such word-learning biases has focused primarily on only one spatial domain to which nouns might refer: objects. In a noun-extension task, we evaluate young children’s and adults’ word-learning biases across two spatial domains—objects and places—and show that young children and adults preferentially extend novel nouns to objects, not the places that the objects are situated in.

Young Children and Adults Extend Novel Nouns to Objects not Places

Systems of semantic categories vary across languages, and it has been proposed that this variation is constrained by a need for efficiency in communication. An important element of efficiency is communicative need, or how often a particular object needs to be referenced. Previous work has sometimes assumed for simplicity that the distribution of need over objects in a semantic domain does not vary across languages or cultures. Here, we explore culture-specific need as it relates to the kinship terminologies of Hindi and English. We assess the efficiency of each language's kin naming system under a variety of need distributions, including one based on that language's usage statistics, one based on the other language's usage statistics, and random permutations of each of those two distributions. Our results suggest that kinship terminologies reflect culture-specific communicative need.

Kinship terminologies reflect culture-specific communicative need: Evidence from Hindi and English

Monolingual and bilingual speakers of the same languages derive unique meanings for words, partly based on between-language differences in meaning of translation approximate words. Do language models also capture this semantic variability? We compared several language models of lexical semantic representation and their correspondences to a behavioral word meaning task done by both L1 and L2 English speakers. We found most language models suggest more similarity between L1 and L2 English speakers and more cross-language similarity between Mandarin and English language models than is supported by psycholinguistic research. We conducted follow-up analysis for the GloVe and OpenAI’s Davinci models and found that individual bilingual participants’ similarity to these models did not relate to language history variables that predict bilingual lexical semantic native-likeness. We conclude that language models are not yet reliable reference points for tracking lexical semantic learning, and we discuss the implications for future directions in computational and psycholinguistics.

Language Models Show Within- and Cross-language Similarities in Concrete Noun Meaning, but not Differences Between L1 and L2 English Speakers

By 24 months, children rapidly infer the meanings of words from sentence context: for instance, inferring from “She eats the dax” that “dax” refers to a food item. However, prior work has typically presented these informative sentence contexts in the presence of the target referent; it is less clear whether sentence context on one exposure will influence learners’ referential hypotheses on subsequent exposures. Here, we present results suggesting that even adults have some difficulty using a word’s prior sentence context to constrain its subsequent hypothesized referents, and 3-year-olds fail to make these cross-exposure inferences at all. Finally, we examine an alternative avenue for sentence context to guide word learning: adults successfully use informative sentence contexts to disambiguate between previously encountered referents. Ongoing work examines this question in children. These results suggest linguistic context is most useful for forming in-the-moment hypotheses but may, in some contexts, also facilitate learning across exposures.

Premium content

Downloads

Next from CogSci 2023

Using measures of metacognition to assess the acquisition of implicit and explicit knowledge in artificial grammar learning tasks

Similar lecture

Leveraging Human Prior Knowledge to Learn Sense Representations