Idiom meaning selection following a prior context: eye movement evidence of L1 direct retrieval and L2 compositional assembly

ABSTRACT

Past work has suggested that L1 readers retrieve idioms (i.e., spill the tea) directly vs. matched literal controls (drink the tea) following unbiased contexts, whereas L2 readers process idioms more compositionally. However, it is unclear whether this occurs when a figuratively or literally biased context precedes idioms. We tested this in an eye-tracking study in which 40 English-L1 and 35 English-L2 adults read English sentences containing idioms having figurative, literal, or control prior contexts. Linear mixed-effects models revealed that L1 readers processed idioms faster after a literal preamble; however, at the disambiguation region, they processed idioms’ figurative interpretations more quickly as familiarity increased, suggesting a L1 reliance on direct retrieval. In contrast, L2 readers processed idioms’ figurative interpretations faster as verb decomposability increased, suggesting an L2 reliance on compositional assembly. Collectively, these results suggest that meaning selection occurs in a hybrid fashion when idioms follow a biased context.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Availability of data and materials

Anonymized data, materials and model summaries can be downloaded from the OSF repository (https://osf.io/6tyn5/).

Code availability

R scripts for models can be downloaded from the OSF repository (https://osf.io/6tyn5/).

Open practices statement

The entire set of anonymized data and materials is available on the OSF repository (https://osf.io/6tyn5/).

Notes

1. Using self-paced reading, Cacciari and Corradini (2015) found contextual effects based on predictability, which is defined as the probability of completing an incomplete phrase idiomatically (e.g., thumbs is more predictable given twiddle your … with respect to bucket given kick the …). Specifically, figurative continuations were read faster after predictable idioms, whereas literal continuations were read faster after unpredictable idioms.

2. Converging findings are found in self-paced reading experiments. Conklin and Schmitt (2008), for instance, reported that both L1 speakers and proficient L2 speakers read figurative and literal uses of idioms comparably when they occurred at the end of a biasing paragraph. Testing sentences with a similar syntactic structure to our stimuli (see below), Beck and Weber (2020) presented L1 readers with idioms preceded by a noun phrase and a relative clause promoting a figurative reading or a literal reading. Both types of sentences could end with a figurative or a literal resolution. Context effects did not emerge in the idiom region, consistent with most of the work reviewed so far, and only involved the resolution region, where literal preambles inhibited the reading of figurative resolutions. Holsinger and Kaiser (2013) obtained parallel results in the disambiguation region for ambiguous verb–particle constructions, and also registered a marginal advantage in the verb region when preambles were figurative.

3. One-way ANOVAs confirmed that sentence preambles did not differ in character length [F(3,153) = 1.77, p > .05] or average log-transformed word frequency [F(3,153) = 1.68, p > .05] across the four conditions. Similarly, resolution regions did not differ in character length [F(3,153) = .02, p > .05] or average log-transformed word frequency [F(3,153) = .60, p > .05]. Word frequencies were taken from SUBTLEX_US (Brysbaert & New, 2009). To make sure that the strength of contextual bias did not differ significantly across conditions, we used pretrained Neural Language Models to compute the probability of idiom verbs and idiom nouns across all conditions. Verb and noun probabilities were extracted from both GPT-2 Large (Radford et al., 2019), an autoregressive model trained on 8 million English web pages, where the probability of the target word is computed based on the left context only, and BERT Large Uncased (Devlin et al., 2019), a bidirectional model trained on 11,038 unpublished books and English Wikipedia, where words are predicted by looking at both the left and right context (Masked Language Modeling). When computing BERT scores, we had to leave out the three idioms buried the hatchet, hedged her bets, and straddled the fence as the words hatchet, hedged, and straddled were not contained in the model. Crucially, one-way ANOVAs failed to detect any significant difference in verb [BERT: F(3,204) = 1.31, p > .05; GPT-2: F(3,216) = .73, p > .05] and noun [BERT: F(3,204) = .94, p > .05; GPT-2: F(3,216) = 1.53, p > .05] context-dependent probabilities between the four experimental conditions. We can tentatively interpret these findings as suggestive that the predictability of idiom forms did not change across sentence types, and thus that the observed differences in reading times reported later on in the article were, rather, due to the different context-based interpretations of the idiom phrases. Finally, we were interested in deriving a readability index for our stimuli. Previous NLP contributions on the topic have observed that relying on lexical features like age of acquisition can provide a reliable readability estimate for single sentences (Schumacher et al., 2016). Accordingly, we took AoA values for each content word in our dataset from Kuperman et al. (2012) and computed an average AoA score for each sentence. A one-way ANOVA confirmed that average AoA did not significantly differ across conditions [F(3,216) = .26, p > .05].

4. Determiner+Noun First Pass Gaze Duration and Idiom First Pass Gaze Duration index the sum of all the fixations made on the determiner+noun and on the idiom region before leaving the region either to the left or to the right. Idiom Total Reading Time is the total duration of all the fixations made on the idiom region, including regressions.

5. To summarize the fixed-effects structure of core models, we predicted log-transformed Det+Noun FPGD and Idiom FPGD from scaled trial order, scaled literal plausibility, preamble (idiomatic vs. literal vs. control, with control or idiomatic as the baseline), language group (L1 vs. L2) and the two-way interaction between preamble and language group. Log-transformed Idiom TRT was instead predicted from scaled trial order, scaled literal plausibility, preamble (biased vs. control), resolution (idiomatic vs. literal), language group (L1 vs. L2), and the three-way interaction between preamble, resolution, and language group.

6. To summarize the fixed-effects structure of L1- and L2-specific models, for L1 and L2 data separately we predicted log-transformed Det+Noun FPGD and Idiom FPGD from scaled trial order, scaled literal plausibility, preamble (idiomatic vs. literal vs. control, with control or idiomatic as the baseline), scaled familiarity, scaled verb relatedness, scaled noun relatedness, the two-way interaction between preamble and scaled familiarity, the two-way interaction between preamble and scaled verb relatedness, and the two-way interaction between preamble and scaled noun relatedness. Log-transformed Idiom TRT was instead predicted from scaled trial order, scaled literal plausibility, preamble (biased vs. control), resolution (idiomatic vs. literal), scaled familiarity, scaled verb relatedness, scaled noun relatedness, the three-way interaction between preamble, resolution and scaled familiarity, the three-way interaction between preamble, resolution and scaled verb relatedness, and the three-way interaction between preamble, resolution, and scaled noun relatedness. For L2 data only, we ran an additional set of models in which scaled familiarity was replaced with scaled cross-language overlap.

7. Regarding the effects of control variables, we found significant negative effects of trial order and literal plausibility except in L2 drill-down models, in which literal plausibility was not significant. Thus, L1 and L2 readers became faster the further they proceeded through the experiment, and L1 readers were generally faster at reading more literally plausible idioms.

8. Accordingly, semantic features of idiom verbs like telicity seem to transfer to the figurative semantics of the whole phrases, and the low frequency of certain idiom verbs (e.g., “twiddle” in twiddle your thumbs or “straddle” in straddle the fence) can act as a strong clue of idiomaticity for comprehenders.

9. Differently from silent reading, experimental paradigms used in other contributions required participants to make more overt semantic judgments on idiom phrases. For example, Gibbs (1980) and Mueller and Gibbs (1987) found L1 speakers to be faster at judging the paraphrases of idioms occurring after a figurative rather than a literal context, but to be more efficient at recalling literal uses of idioms. When idiom meaning selection was tested with cross-modal priming in a minimal context, initial evidence suggested that L1 speakers activated both literal and figurative meanings over different time courses and depending on idioms’ predictability (Cacciari & Tabossi, 1988), whereas L2 speakers were characterized by an overall literal bias (Cieślicka, 2006). In light of this literal bias emerging in L2, Cieślicka (2006, 2010) drew on the graded salience hypothesis by Giora (1997, 1999, 2003) to propose a literal salience resonant model for L2 idiom processing. Just like the graded salience hypothesis predicted a processing facilitation for the meaning (literal or figurative) that is more frequent and conventional in the speakers’ lexicon, according to Cieślicka’s model literal meanings of L2 idioms enjoy a more salient status, even after L2 idioms are incorporated into the learners’ lexicon, as learners are likely to use and come across literal meanings more often. Nevertheless, more recent data coming from cross-modal and visual-visual paradigms showed that L1 and L2 speakers were similarly facilitated by both figurative and literal targets over unrelated targets (Beck & Weber, 2016; Van Ginkel & Dijkstra, 2020). With an interesting manipulation that tried to bridge the gap between reading and priming studies, Kyriacou et al. (2022) used eye-tracking during reading to analyze the downstream activation or figuratively and literally related words after reading passivized idioms at the beginning of a sentence. Literally related targets seemed to be activated faster here, probably because idioms were encountered in a less familiar, passivized form.

Additional information

Funding

This research was supported by the Natural Science and Engineering Research Council of Canada (RGPIN-2022-03375 to DT).