Limits on the Agent‐First Strategy: Evidence from Children's Comprehension of a Transitive Construction in Korean

Abstract

It has long been believed across languages that the Agent‐First strategy, a comprehension heuristic that maps the first noun onto the agent role, is a general cognitive bias which applies automatically and faithfully to children's comprehension. The present study asks how this strategy interplays with such grammatical cues as the number of overt arguments and the presence of case‐marking in Korean, an SOV language with case‐marking by dedicated markers. To investigate whether and how these cues affect the operation of this strategy, we measure children's comprehension of a transitive construction (with scrambling and omission of sentential components) in a novel experimental setting where arguments and case markers were obscured to varying degrees through acoustic masking. We find that children do not demonstrate the agent‐first interpretation strongly in the noun–verb pattern without case‐marking, showing their uncertainty about the thematic role of the nominal when it is both the only argument in the sentence and lacks case‐marking. They perform significantly better in the patterns with additional cues, the impact of which is asymmetric by age and by the nature of alignment between cues from word order and case‐marking. These findings suggest that, for Korean‐speaking children's comprehension of a transitive construction, the Agent‐First strategy is activated properly only in conjunction with other types of interpretive cues.

Explainability of neural networks for child language: Agent‐First strategy in comprehension of Korean active transitive construction

“…For test items, we employed the same stimuli used in Shin (2021). Each condition consisted of six instances, with animals as agents and themes and actional verbs at the end (Table 3).…”

“…However, Shin (2021) reveals some limits on this bias, arguing that there may be no standalone Agent‐First strategy for comprehension (see also Garcia & Kidd, 2020). Shin finds that, for Korean‐speaking children's comprehension of a transitive event, the Agent‐First strategy is activated properly only in conjunction with other types of grammatical cues.…”

“…Shin (2021) notes four major findings (Table 1). First, whereas the children had a good command of case‐marking knowledge regarding the active transitive (the nominative case marker indicating the agent; the accusative case marker indicating the theme), they showed an asymmetry in performance by canonicity: they were better in the canonical condition (N NOM N ACC V) than in the scrambled condition (N ACC N NOM V).…”

“…Based on these findings, Shin (2021) argues that, when Korean‐speaking children interpret a transitive event, they do not employ this strategy automatically and immediately based solely on an argument's initial position in the sentence. Considering the particular experimental setting in which participants were exposed to pictures prior to stimuli so that they adjust their interpretation to transitive events with two animate entities (one as an agent and the other as a theme) before encountering the stimuli, the children's comprehension behaviour would have been guided by two major forces.…”

“…We investigate whether and how NNs, as a proxy for cognitive space wherein learning occurs, reveal children's manifestation of the Agent‐First strategy in comprehension. We develop four NN models— Word2Vec (Mikolov et al., 2013), Long Short‐Term Memory ( LSTM ; Hochreiter & Schmidhuber, 1997), Bidirectional Encoder Representations from Transformers ( BERT ; Devlin et al., 2018), Generative Pre‐trained Transformer 2 ( GPT‐2 ; Radford et al., 2019)—and measure their classification performance on the same stimuli used in Shin (2021). Given the special status of this strategy in child language development as a window to the interface between linguistic knowledge and domain‐general factors, scrutinising the extent to which deep‐learning algorithms capture children's language behaviour with respect to this comprehension bias is expected to reveal the explainability of artificial intelligence for child language, and more fundamentally, for (the developing nature of) a child processor.…”