Learning about the meanings of ambiguous words: evidence from a word-meaning priming paradigm with short narratives

Introduction

Language is highly ambiguous. As over 80% of common English words have multiple dictionary definitions (Rodd, Gaskell & Marslen-Wilson, 2002), the ability to rapidly select context-appropriate word meanings is a key component of skilled comprehension. Disambiguation processes are particularly complex for subordinate (less frequently used) word meanings, e.g., when the word “crane” refers to a bird rather than to a tall metal structure on a building site. People spend longer reading ambiguous words that are disambiguated towards their subordinate meaning, compared with unambiguous words and dominant meanings (e.g., Duffy, Morris & Rayner, 1988; Pacht & Rayner, 1993). In spoken language, a similar increase in processing time and resources for ambiguous compared to unambiguous words has been shown in studies using a range of approaches including neuroimaging (e.g., Rodd, Davis & Johnsrude, 2005; see Vitello, 2014 for review), dual-task methods (Rodd, Johnsrude & Davis, 2010), picture selection tasks (Foss, Bever & Silver, 1968) and pupillometry (Kadem et al., 2020).

Recent word-meaning priming experiments suggest that learning processes play a key role in word-meaning disambiguation. Following successful disambiguation of an ambiguous word, people retain information about the encounter. The availability of the selected meaning is temporarily boosted for easier access at subsequent encounters of the same word form (Betts et al., 2018; Gaskell, Cairney & Rodd, 2019; Gilbert et al., 2018; Gilbert et al., 2021; Rodd, Johnsrude & Davis, 2010; Rodd et al., 2013; Rodd et al., 2016; see Rodd, 2020 for review). This temporary availability boost represents a form of lexical learning that aids language interactions by supporting access to relatively rare word meanings that are used consistently within texts or conversations. Similarly, work on text reading has found that shifting between the different meanings of an ambiguous word is associated with processing costs (i.e., longer reading times, Binder & Morris, 1995; Binder & Morris, 2011). More broadly, these findings demonstrate people’s sensitivity to their current linguistic environment and the flexibility of the adult lexical processing system (Gaskell, Cairney & Rodd, 2019).

The word-meaning priming effect is well replicated. Experiments typically consist of a priming phase in which ambiguous words are strongly disambiguated by their sentence context followed by a test phase. Tasks such as word association (Betts et al., 2018; Gaskell, Cairney & Rodd, 2019; Gilbert et al., 2018; Gilbert et al., 2021; Rodd et al., 2013; Rodd et al., 2016) or speeded semantic relatedness judgement (Gilbert et al., 2018; Gilbert et al., 2021) are used to tap word meaning availability at test. One limitation of these experiments is that the priming phase typically presents the ambiguous words within lists of unconnected sentences that contain strong local disambiguation cues (although see Betts et al., 2018; Rodd et al., 2016 for experiments with longer paragraphs). In natural language, by contrast, disambiguation may also occur based on more distant contextual cues embedded in a longer communicative context (but see Witzel & Forster, 2015; Rice et al., 2019 for discussion of the importance of local lexical cues in disambiguation). The effect of global (rather than sentence-internal, local) context has previously been investigated during reading (e.g., Binder, 2003; Kambe, Rayner & Duffy, 2001). Irrespective of whether context occurred locally or globally, reading times were longer for ambiguous words when the context was consistent with the word’s subordinate (rather than dominant) meaning. The finding that this subordinate-bias effect (Pacht & Rayner, 1993) is similar across both local and global contexts shows that readers are able to use more distant contextual cues to disambiguate word meanings further downstream.

In the present work, we explored listeners’ ability to process and learn from short narratives where disambiguation was driven by global rather than local context (see examples in Table 1). The priming phase comprised three-sentence narratives that ended with an ambiguous word prime (e.g., “crane”). The first sentence of each narrative established a situational context congruent with the subordinate meaning of the prime, but the remainder of the narrative provided no disambiguating information. During this priming phase, participants completed a semantic relatedness task in which they responded to visually presented word probes to ensure they were listening for meaning. After a brief delay filled by a digit span task lexical learning was assessed in the form of a word association task. This test task probed the impact of the contextual exposure to the ambiguous word during the priming phase on a subsequent, context-free encounter (i.e., word association task) of the same word forms (Gilbert et al., 2018; Twilley et al., 1994).

In addition to lexical learning, the word-meaning priming paradigm allowed us to assess the success of disambiguation processes: only if participants have correctly selected the subordinate meaning during the priming phase can the availability of this specific meaning receive a boost. While eye-movement studies have provided evidence that people can use distant contextual cues for disambiguation when reading (e.g., Hess, Foss & Carroll, 1995; Kambe, Rayner & Duffy, 2001; Schwanenflugel & White, 1991), it has proven more difficult to measure equivalent disambiguation processes for spoken language (see Rodd, Johnsrude & Davis, 2010). Our paradigm provides an indirect approach to measuring disambiguation in spoken language, and one that has clear advantages over tasks that measure comprehension more directly e.g., via explicit comprehension questions, or by asking people to make semantic relatedness judgements (e.g., CRANE—BIRD) that may themselves change behaviour by explicitly priming particular meanings. Using a priming approach with spoken language, Gilbert et al. (2021) have previously replicated key findings from the reading literature. In the priming phase of their experiment, disambiguating information congruent with the subordinate meaning occurred either before or after the ambiguous word. Priming was reduced when disambiguation cues occurred after the ambiguous word, suggesting that listeners were less likely to successfully resolve the ambiguity in late-disambiguation compared to early-disambiguation sentences, in line with findings from the literature on sentence reading (e.g., Frazier & Rayner, 1990). Priming was also reduced when the priming task could be completed via relatively shallow processing (listening to sentences with only occasional comprehension questions vs. making a semantic relatedness judgement after every sentence). This is in line with the “good enough” account of sentence processing (e.g., Ferreira & Patson, 2007), whereby readers tend to process complex language in a way that is “good enough” for the task at hand - for example, by not always resolving syntactic (Christianson et al., 2001) or semantic ambiguities (Blott et al., 2021). In their word-meaning priming study, Gilbert et al. (2021) found that, like readers, listeners do not always fully engage in disambiguation processes. The presence of a priming effect in the current experiment will therefore provide insight into the likelihood of disambiguation success when spoken contextual cues to disambiguation are relatively distant from the ambiguous word itself, potentially replicating findings from the literature on written sentence processing (e.g., Hess, Foss & Carroll, 1995; Kambe, Rayner & Duffy, 2001; Schwanenflugel & White, 1991)

We hypothesised that subordinate meanings would receive an availability boost when they had recently been encountered (Primed condition) compared to when they had not (Unprimed condition, see preregistration). To preface an exploratory analysis without a specific hypothesis, we also investigated whether priming is modulated by word meaning dominance. Rodd et al. (2013, Exp. 1) found greater priming for subordinate meanings of words that were strongly biased (e.g., critter vs listening device “bug”) compared to words whose meanings are relatively equal in frequency (e.g., mammal vs container “seal”). Plausibly, this finding reflects the fact that a single encounter with a meaning is highly informative when the described event is unexpected—as is the case for more strongly subordinate meanings. However, it is also plausible that strongly subordinate meanings might be less reliably disambiguated, resulting in reduced priming for these items. Any effect of dominance on word-meaning priming (i.e., an interaction of Dominance × Priming) in the present data may help to understand dominance within theoretical accounts of word-meaning priming (e.g., by taking into account baseline dominance as a limit to possible shifts in meaning availability through priming, Rodd, 2020), and guide stimulus selection in future studies.

In summary, the current experiment assessed whether participants can disambiguate spoken ambiguities on the basis of distant contextual cues, whether these disambiguation encounters are used for lexical learning and affect word meaning interpretation at a later time point, and to what extent such priming effects are modulated by word meaning dominance.

Materials & Methods

Hypotheses, methods and data analysis plan were preregistered, with deviations noted where necessary: https://osf.io/xwn5h. Data and analysis code are available here: https://osf.io/bvnhk/.

Procedure

The experiment was programmed using Gorilla experiment builder (http://www.gorilla.sc/about; Cauldron Inc.; Anwyl-Irvine et al., 2019), and completed online. Participants used their own computer (the experiment was not accessible via a phone/tablet) and were instructed to wear headphones. The experiment comprised six phases (Fig. 1), which were not separated by enforced breaks between them—instead, participants automatically moved from one task to the next.

Instructions to enable the browser “autoplay sounds” option were provided in the audio-check phase (Milne et al., 2021). Participants were reminded to wear headphones before completing six trials of a Huggins Pitch task, which includes tones that are only audible through headphones (Milne et al., 2021). Participants then listened to an example sentence and adjusted computer volume levels accordingly.

In the priming phase, each participant listened to 33 narratives that ended in an ambiguous prime, and 33 unambiguous fillers (items were counterbalanced across participants, see Materials for details). Trial order was randomised for each participant. To encourage listening for meaning, they made simple semantic relatedness decisions to visually presented probe words at the offset of each spoken narrative (responding ‘c’ for related, or ‘m’ for unrelated probes). In total, 50% of trials were followed by a related probe. Responses did not time out. The experiment began with four practice trials (three unambiguous fillers and one ambiguous narrative similar to the experimental stimuli). Participants received accuracy feedback for practice trials only. The main task started with four additional unambiguous filler narratives to allow participants to become accustomed to the trial structure. Attention was not drawn to the presence of ambiguous words and participants were not told that some words would appear again later in the test phase.

A forward and backward digit span task from the Gorilla Open Materials depository (Dean, 2020) provided a delay between the priming and the test phases (M = 6.86 min, SD = 1.86, Range = 4–13 min).

In the test phase all 66 ambiguous words were presented aurally in a randomised order for word association. Participants were instructed to type the first word that came to mind. Each trial timed out after 30 s. The mean difference between the estimated average mid-point of the priming phase and the mid-point of the test phase was approximately 20 min, which is in line with previous studies on word meaning priming (Betts et al., 2018; Gilbert et al., 2021; Rodd et al., 2013).

Finally, participants were asked two open-ended questions in the debrief questionnaire: “Did you notice anything about the stories or words that were presented to you?” and “What do you think the aim of this study was?”.

Results

Test phase: word association task

Each response was coded as either consistent with the subordinate meaning used in the context of the priming phase (coded as ‘1’) or inconsistent (coded as ‘0’). Responses were independently coded by two authors, blind to priming condition. We excluded responses that were related to both meanings or likely reflected mishearings or were otherwise unclear. The coders initially agreed on 97.3% of decisions. Disagreements were resolved by discussion. 332 trials (8.98%) were excluded.

As per the preregistration, the data were analysed with logit mixed effects models (Baayen, Davidson & Bates, 2008; Barr et al., 2013), using the glmer() function (with bobyqa optimiser) within the lme4 package (v1.1.26; Bates et al., 2015) in RStudio (v3.6.2; RStudio Team, 2015). The factor “Priming” was deviation-coded (Primed 0.5, Unprimed −0.5). Our preregistration specified that we would include a fixed effect for stimulus list; however, the two factors of Priming and List were redundant, so we excluded the fixed effect for List. Although not specified in the preregistration, our maximal model also included random slopes for Priming to account for inter-individual and inter-item differences in the effect of priming on word association responses. Due to convergence issues, we reduced the complexity of the random effects structure (following Barr et al., 2013); the maximal model included a fixed effect for Priming and random intercepts by participants and items.

As predicted, responses were more likely to be consistent with the subordinate meaning if participants had encountered that meaning in the Priming Phase (Fig. 2; b = 0.53, SE = 0.10, z = 5.13; χ²(1) = 25.77, p < .001).

For an exploratory analysis of how a word’s relative meaning dominance affected priming, we used a measure of dominance previously calculated from an independent data set of word association responses (Gilbert & Rodd, 2022). This measure captures the likelihood with which people produce responses associated with the meanings of an ambiguous word, as a value between 0 (both meanings equally likely) and 1 (one meaning produced exclusively). We fitted a model with a continuous fixed effect for the item’s dominance (D as calculated in Gilbert & Rodd, 2022), which was centred before inclusion in the model, a fixed effect of Priming and their interaction. Due to convergence issues, we reduced the complexity of the random effects structure following recommendations in Barr et al. (2013). The maximal model that converged included random intercepts by participants and by items. This analysis found a significant main effect of Dominance; compared to items with lower Dominance values, items with higher baseline Dominance were associated with fewer responses consistent with the subordinate meaning (b = −5.63, SE = 0.58, z =  − 9.77; χ²(1) = 66.1, p < .001). This negative relationship between baseline dominance and consistency of word association responses is illustrated in Fig. 3. It confirms that the participants in this experiment had similar meaning preferences to those sampled by Gilbert & Rodd (2022). However, in contrast with our prediction, Dominance did not interact with Priming (see Fig. 4); there was no evidence to suggest that the magnitude of priming was dependent on the baseline dominance of the item (b = 0.59, SE = 0.44, z = 1.34; χ²(1) = 1.75, p = .186). Importantly, the main effect of Priming on consistency remained significant when baseline dominance was included in the model (b = 0.63, SE = 0.13, z = 5.07; χ²(1) = 25.41, p < .001).

Discussion

This study found that a single encounter with the subordinate meaning of an ambiguous word within a spoken narrative increased the availability of this meaning at a subsequent encounter, replicating the word-meaning priming effect (Betts et al., 2018; Gaskell, Cairney & Rodd, 2019; Gilbert et al., 2018; Gilbert et al., 2021; Rodd et al., 2013; Rodd et al., 2016). Notably, our experiment differed from previous work as ambiguous words occurred within narratives rather than in a list of unconnected sentences, and disambiguation was driven by contextual cues separated from the ambiguous word by an intervening sentence, rather than by sentence-internal lexical cues. Participants showed evidence of lexical learning in the present experiment. That is, they retained information about the meanings they had encountered, and this guided their subsequent interpretation of the words. We provide evidence that lexical learning can occur through naturalistic comprehension processes, even when successful comprehension relies on the use of global contextual cues to disambiguate strongly subordinate word meanings. The present work adds to the evidence that word-meaning priming is a robust phenomenon (Betts et al., 2018; Gilbert et al., 2021; Rodd et al., 2013; Rodd et al., 2016). Indeed, priming effects have now been observed when the priming phase required disambiguation based on local sentence-internal disambiguation cues occurring before (Rodd et al., 2013; Rodd et al., 2016) or after the ambiguous word itself (Gilbert et al., 2021), and based on distant disambiguation cues from earlier in a paragraph (Betts et al., 2018) or narrative (present study).

Although the availability of subordinate meanings was boosted by priming, the baseline dominant meanings were still preferred in the word association task overall. This is a consistent finding in the word-meaning priming literature, with the average proportion of word association responses consistent with subordinate meanings usually not exceeding about 0.35 (Gilbert et al., 2018; Gilbert et al., 2021; Rodd et al., 2013; Rodd et al., 2016), and even when the subordinate meaning was primed multiple times or prime encounters were spaced across time (Betts et al., 2018). This suggests that prolonged and consistent changes in the linguistic environment may be needed to affect larger changes to the baseline meaning availability that has accumulated through long-term experience across the lifespan (cf Nation, 2017). This type of change might be induced by a change in an individual’s hobbies or profession, for example (Eligio & Kaschak, 2020; Rodd et al., 2016; Wiley, George & Rayner, 2018).

Baseline dominance had a clear effect on the availability of the subordinate meanings within the experiment: the less strongly an ambiguous word was biased towards a dominant meaning, the more likely participants were to respond with the subordinate meaning. In contrast to Rodd et al. (2013) however, and despite including ambiguous words with a relatively wide range of dominance values, there was no evidence to suggest that baseline dominance modulated the size of the priming effect. This null finding should be interpreted with caution but we speculate that for the type of contextual disambiguation required in the current study, the magnitude of priming may be more strongly influenced by other item-specific factors, such as the strength of the disambiguating context (see e.g., Colbert-Getz & Cook, 2013; Kambe, Rayner & Duffy, 2001; Martin et al., 1999 for evidence that contextual strength can modulate disambiguation processes during reading). Such factors may have masked any effects of baseline dominance. Further studies that explicitly manipulate word-meaning dominance and contextual strength are needed to disentangle the factors that drive successful disambiguation in spoken language comprehension, and the subsequent effects of lexical learning from encounters with ambiguous word meanings and how they build up over time.

The nature of the cognitive mechanisms that support the short-term shifts in meanings preferences induced by word-meaning priming is still debated. While the present study cannot answer questions about the underlying mechanisms of these priming effects, it will be important for future work to distinguish between accounts that see the origin of word-meaning priming in an immediate alteration of long-term memory (Rodd et al., 2016; Gilbert et al., 2018), and those that see the priming phenomenon as the result of a reinstantiation during the testing phase of a gist-like episodic memory representation of the word-meaning encounter during the priming phase (Curtis et al., 2022; Gaskell, Cairney & Rodd, 2019). Our findings show that successful disambiguation for spoken ambiguities can occur based on contextual representations built earlier in the discourse. This result complements findings from eye-tracking studies in the written modality (e.g., Hess, Foss & Carroll, 1995; Kambe, Rayner & Duffy, 2001) and extends them to spoken language comprehension, using an implicit measure of disambiguation success.

Conclusions

This experiment is an important step towards using more naturalistic and varied forms of disambiguation to study how people use cues from recent experience to update their lexical knowledge in a manner that supports comprehension. It contributes to the growing literature demonstrating the flexibility of the adult lexical processing system (Rodd, 2020) and shows the importance of linguistic experience across the lifespan in maintaining high quality lexical representations (Nation, 2017).

Supplemental Information