Is Chunking Ability Unitary Across Levels of Linguistic Representation?

Abstract

Much like with statistical learning, there is a tendency in the cognitive and language science literature to treat chunking as a unitary, domain‐general ability that operates fairly similarly across domains of cognition. An alternative position suggests chunking in language and cognition involves multiple mechanisms tuned to the specific perceptual properties and statistical regularities of the target structures of learning. To test these hypotheses, we compared performance on two different forms of chunking in English: (1) multiword chunking as assessed by a recall task involving multiple high‐ and low‐frequency three‐word sequences, and (2) clausal chunking in which participants intuitively insert chunk boundaries on a touchscreen tablet while listening to speech extracts and viewing the corresponding transcripts. Results indicated no correlations between multiword chunking and clausal chunking, providing initial support for the notion that chunking processes may likely not operate equally across or even within the same cognitive domain, casting doubt on a unitary view of chunking. Despite limitations in the tasks used to assess chunking, they nonetheless provide a useful starting point for future studies looking into the nature of chunking across different levels of linguistic representation.

1. Introduction

There has been a resurgence of interest in chunking as a key mechanism for learning and processing statistical regularities in language across multiple levels of linguistic representation, from syllables and words to phrases and large syntactic patterns (e.g., Christiansen & Chater, ²⁰¹⁶; de la Cruz‐Pavía, Werker, Vatikiotis‐Bateson, & Gervain, ²⁰²⁰; MacWhinney, ²⁰⁰⁵; Mauranen, ²⁰⁰⁹). Chunking, defined here as an experience‐dependent ability to group linguistic elements into larger units (Gobet et al., 2001), allows language users to deal more efficiently with the deluge of input characteristic of everyday linguistic interactions (Christiansen & Chater, 2016). Although chunking is typically considered to be domain‐general in nature (Gobet et al., 2001), the question remains as to whether chunking constitutes a unitary psychological construct that functions in the same way across different domains, or whether multiple distinct chunking abilities exist, each tuned to the specifics of a particular domain (or aspect thereof). In this article, we take a first step toward addressing this question empirically, by studying the relationship between two different linguistic chunking tasks aimed at capturing chunking at different levels of linguistic representation.

1.1. Unitary versus multiple‐mechanisms view of chunking

The question of a single, unitary mechanism versus multiple mechanisms for chunking has a parallel in the related study of statistical learning (SL). Researchers have long operated under the assumption of a general SL ability employed in learning statistical regularities across different domains of cognition (e.g., Bulf, Johnson, & Valenza, ²⁰¹¹; Endress & Mehler, ²⁰⁰⁹; Kirkham, Slemmer, & Johnson, ²⁰⁰²; Parks, Griffith, Armstrong, & Stevenson, ²⁰²⁰; Saffran & Thiessen, ²⁰⁰⁷). For example, the alternating serial reaction time task, the tower of London/Hanoi task, and the weather prediction task have been tested as correlates of different aspects of language processing (Brill‐Schuetz & Morgan‐Short, 2014; Faretta‐Stutenberg & Morgan‐Short, ²⁰¹⁸; Pili‐Moss, Brill‐Schuetz, Faretta‐Stutenberg, & Morgan‐Short, ²⁰²⁰). The typical (often implicit) assumption is that SL is subserved by a unitary mechanism, with individual differences in SL ability in one domain (e.g., language) expected to relate to SL ability in other domains (e.g., object perception); moreover, SL ability should also be similar within the same cognitive domains, for example, SL at the phonological and multiword levels in language processing. Thus, according to this view, a single unitary mechanism for learning serves in the acquisition of statistical regularities across domains regardless of the target structures of learning or their input characteristics.

In contrast, a multiple‐mechanisms view suggests that while SL processes are engaged in different domains of cognition, they may operate differently across and even within domains based on the properties of the regularities acquired (Bogaerts, Siegelman, Christiansen, & Frost, 2022; Frost, Armstrong, Siegelman, & Christiansen, ^{2015, 2019}). Under this multiple‐mechanism view, SL occurs with respect to specific abilities (e.g., reading, syntax, scene perception) that are fine‐tuned to the particular properties of the target domain of learning. Supporting this view, weak or nonexisting correlations in SL performance have been observed for learning different kinds of statistical structure within a given modality (e.g., Misyak & Christiansen, ²⁰¹²; Siegelman & Frost, ²⁰¹⁵) and for learning the same kind of structure across modalities (e.g., Arnon, ²⁰²⁰; Siegelman & Frost, ²⁰¹⁵), suggesting that multiple distinct computations may underlie learning in different domains. If SL ability indeed operates in accordance with a multiple‐mechanisms view, then chunking in language may also reflect such domain‐specificity optimized for representations based on, for example, temporal, structural, and combinatorial input characteristics (Christiansen, 2019). Thus, we hypothesize that chunking in language is subject to the same constraints that give rise to variability across modalities in SL.

As our theoretical approach to chunking in language, we adopt the chunk‐and‐pass processing framework (Christiansen & Chater, 2016), which suggests that chunking occurs separately at different levels of linguistic representation. In language comprehension, the framework proposes that the speech signal is initially chunked into phonemes or syllables and passed up to higher levels of representation to form words or multiword chunks, and so on to achieve discourse comprehension. In language production, the process is reversed with the discourse‐level message initially broken down into intonational units, which are then split into multiword sequences and words that are subsequently divided into syllables or phonemes before being turned into the motor commands necessary to articulate such sounds. In support of this perspective, McCauley, Isbilen, and Christiansen (2017) found that chunking at the phonological and multiword levels did not correlate, suggesting that distinct mechanisms may underlie chunking even within the same domain.

To further test the multiple‐mechanism hypothesis, we employ two existing linguistic chunking tasks that are hypothesized to tap into different levels of chunking: multiword combinations and clauses. The multiword chunking (MWC) recall task aims to capture linguistic regularities at the level of three‐word sequences from exposure to natural language statistics (e.g., Isbilen, McCauley, & Christiansen, ²⁰²²). Sensitivity to such linguistic regularities is thought to facilitate language comprehension (Arnon & Ramscar, 2012; Arnon & Snider, ²⁰¹⁰; Tremblay, Derwing, Libben, & Westbury, ²⁰¹¹) and production (Arnon & Christiansen, 2017; Bannard & Matthews, ²⁰⁰⁸; McCauley & Christiansen, ²⁰¹⁷; Siegelman & Arnon, ²⁰¹⁵). The clausal chunking (CC) task is designed to capture an intuitive capacity to segment the incoming speech stream incrementally into temporal chunks to aid in language comprehension (Sinclair & Mauranen, 2006; Vetchinnikova, Mauranen, & Mikušová, ^{2017, 2022}). We note that our description as “clausal” is different from the authors’ use of “perceptual” chunking, as we believe our description more accurately characterizes the higher‐level syntactic clausal representations captured in these chunks. From this perspective, we are investigating whether chunking at the level of syntactic clauses is associated with chunking at the lower level of multiword sequences. Syntactic regularities are thus involved in both chunking tasks, given that the three‐word chunks used in the MWC recall task all correspond to standard syntactic constituents. Through the lens of the single‐mechanism perspective, individual differences across the two tasks should be correlated regardless of target structures, whereas a weak or nonexisting relationship is expected from the multiple‐mechanisms viewpoint (cf. Bogaerts et al., ²⁰²²; Christiansen & Chater, ²⁰¹⁶).

1.2. The current study

This study aims to compare predictions from the single‐ versusmultiple‐mechanisms accounts regarding the nature of chunking in language, using two previously developed tasks that tap into chunking abilities at different levels of linguistic structure: three‐word sequences and clauses. MWC ability has typically been assessed using recall measures in natural and artificial languages at different levels of representation, for example, at the utterance level in elicited imitation tasks using naturalistic speech (Culbertson, Andersen, & Christiansen, 2020) and at the phonological level in statistically induced chunk recall following exposure to an artificial speech stream (Isbilen, McCauley, Kidd, & Christiansen, 2020, 2022). Additionally, McCauley et al. (2017) provide a recall measure of SL at the multiword level using word trigrams of varying frequency derived from corpus data. Such measures have been interpreted as reflecting the learning of recurrent linguistic regularities that characterize language fluency (Bowden, 2016; Klem et al., ²⁰¹⁵).

CC has been assessed using an online word segmentation task designed to capture intuitively identified chunk boundaries (Vetchinnikova et al., 2017, 2023). The underlying assumption is that participants intuitively segment continuous speech into chunks that can be processed further for meaning (Mauranen, 2025; Sinclair & Mauranen, ²⁰⁰⁶). It is further hypothesized that fluent speakers of a language will mostly converge on the same chunk boundaries that may serve as a basis for describing grammatical structure. Our study further sought to replicate the results from the CC task in Dobrego, Konina, and Mauranen (2023) using the same speech extracts with naïve native English speakers.

Our research question asks if MWC recall ability correlates with CC, in accordance with either a single‐ or multiple‐mechanism view. We predict that under a single‐mechanism view, there will be a positive linear relationship between MWC recall ability and CC and no relationship under a multiple‐mechanism view.¹

2. Methods

2.1. Participants

In this preregistered study: https://aspredicted.org/vhg3‐7vwy.pdf, we sought to recruit 70 native speakers of English at the undergraduate level. This number was based on a power analysis using G*Power version 3.1 (Faul, Erdfelder, Buchner, & Lang, 2009) for a one‐tailed test with a medium effect size of d = 0.3, power = 0.8, and p = .05. We recruited a total of 80 participants, of whom five were excluded due to technical issues or inability to complete the study correctly, and six due to excessive boundary inserts defined as any observation more than 1.5 times the interquartile range above the third quartile following Dobrego et al. (2023).

The final analysis consisted of 69 participants (45 females, 23 males, and 1 other), with ages ranging from 18 to 23, with a mean age of 20.11 years (SD = 1.15). All participants were native English speakers with no uncorrected hearing difficulties (60 were right‐handed). Ten participants were native bilingual speakers reporting a second native language other than English, and most participants (63.76%) reported speaking a second/non‐native language.

3. Materials

3.1. MWC task

For this task (taken from Isbilen et al., ²⁰²²), participants listened to a total of 20 strings, each consisting of 12 words that were recorded separately to avoid any prosodic cues. Half of the strings consisted of four frequent three‐word sequences from natural language, and the other half consisted of the same 12 words pseudorandomized to minimize any bigram/trigram information. Participants listened to each 12‐word string and then typed the words back as best as possible in the order they were presented. Performance was based on individual word recall (number of individual words recalled in the correct order) as well as trigram recall (number of trigrams recalled in any order, given that recalling a trigram still reflects successful chunking and order of trigram recall would likely not affect overall string representations as each string is composed of trigrams with unrelated discontinuous information, unlike recall of full natural discourse).

3.2. CC task

The CC task is an online processing task that assesses the ability to intuitively identify chunk boundaries in speech (Dobrego et al., 2023; Vetchinnikova et al., ²⁰¹⁷). Developed as a touchscreen application, the ChunkitApp is an open‐source tool available at: https://chunkitapp.online. In this task, participants listen to recorded speech while viewing a transcript of it on a touchscreen tablet. They are instructed to insert chunk boundaries by tapping on the spaces between words, each marked with an interactive tilde symbol (“∼”). When tapped, the tilde changes to a boundary (“|”) mark. Participants can remove or reinsert a chunk boundary by tapping on the space again. Recordings can also be paused, which makes the transcript disappear and participants cannot insert boundaries until the recording resumes. Participants were given minimal instructions for completing the task, and no explanation was provided for what constitutes a chunk. The instructions, taken from Dobrego et al. (2023), were as follows:

Humans process information constantly. When we take in information, we tend to break it up quickly into small bits or chunks. We ask you to work intuitively. Your task is to mark boundaries between chunks by clicking ‘∼’ symbols. One click makes the boundary appear. If you click the symbol again, the boundary will disappear. If you lose the line in the text, stay with the speaker and do not try to go back. (p. 494)

The CC task included a total of 50 speech extracts and two practice extracts, all from two corpora: the Michigan Corpus of Academic Spoken English and the English as a Lingua Franca in Academic Settings Corpus. We used the same recordings from Dobrego et al. (2023) in which they re‐recorded the original audio with a fluent speaker of English replicating the intonation patterns of the original speech. Each extract was 25–30 s long and contained a meaningful passage related to a real‐world academic context with few hesitations, proper nouns, and technical terms. There were a total of 2530 potential boundaries, one less than the total number of words across all extracts. Following each recording, participants answered a yes/no comprehension question about the extract to ensure it was processed adequately for meaning. Fig. 1 shows (a) a sample extract as well as (b) the same extract with the proportion of boundary inserts between each word indicated.

Fig. 1.

(a) Sample extract from the ChunkitApp. Participants see a “∼” symbol between each word indicating potential boundaries. If tapped, a boundary (“|”) appears, and if retapped, reverts. (b) The relative frequencies in between each word indicate the proportion of boundaries inserted at a given location. Relative frequencies in bold indicate a significant chunk boundary and are marked with the boundary symbol.

The key dependent variable for this task is whether participants marked a chunk boundary or not for each between‐word space. We assigned a point value for agreement for each potential chunk boundary (i.e., each space between each word) by calculating the proportion of native English speakers from Dobrego et al. (2023) who marked or did not mark each boundary. We used these generated proportions to assess the agreement rating of each of our participants. All code and shareable data are available on the Open Science Framework at: https://osf.io/jy924/overview?view_only=1bcefc542d4c4bbdbe8c66d1cfc3c12d.

3.3. Procedure

Participants completed all tasks in the same order: an informed consent form, a demographics questionnaire, the CC task, and the MWC task.² Moreover, both chunking tasks involved the same pseudorandomized order of items across all participants. We thus followed prior work on individual differences in language processing (e.g., Frinsel & Christiansen, ²⁰²⁴; James, Fraundorf, Lee, & Watson, ²⁰¹⁸; Liceralde et al., ²⁰²⁶; Swets, Desmet, Hambrick, & Ferreira, ²⁰⁰⁷), keeping task and item order constant across participants to help minimize irrelevant between‐participant variance not directly related to differences in their individual ability to chunk clauses and multiword sequences.³

4. Results

4.1. MWC task

All analyses were carried out in R, version 4.4.2 (R Core Team, ²⁰²⁴). A linear mixed‐effects model run using the “lme4” package (Bates, Mächler, Bolker, & Walker, 2015) showed that, for word recall with fixed effects for subject and item number, participants recalled significantly more words from experimental compared to random strings (t(18) = 15.01, SE = 0.30, p < .001). Participants also showed significantly better trigram recall on experimental compared to random strings (t(18) = 13.92, SE = 0.16, p < .001).

Assessing the internal consistency of our MWC difference (experimental‐control) scores using the “psych” package (Revelle, 2024), we found that Cronbach's alpha was acceptable for word (α = 0.78) but not trigram (α = 0.65) recall. To keep the measures of chunk recall consistent in our models, we chose to instead include target word and target trigram recall, both of which showed good levels of reliability (α = 0.88 and α = 0.80, respectively). It is important to note that recall of target strings does not control for baseline working memory, which is captured by random strings and subtracted out in the difference scores. Thus, our interpretation of this measure is limited in scope without regard to differences in baseline working memory among our participants. We also note that this change in response measure diverges from the initial proposed analysis in our preregistration. Summary statistics for both measures of MWC recall are presented in Table 1.

Table 1.

Multiword chunking results

	Words recalled			Trigrams recalled
	Mean	SD	Range	Mean	SD	Range
Target	9.59	2.47	0−12	2.67	1.18	0−4
Random	4.96	1.60	0−12	0.33	0.52	0−4

4.2. CC task

To assess individual differences in rate of agreement among our participants, we scored agreement in relation to the native English‐speaking participants from Dobrego et al. (2023). For each potential boundary (i.e., space between two words), we calculated the relative frequency of the Dobrego participants who did or did not insert a chunk boundary and assigned that point value for each response. For example, if 75% of participants marked a certain boundary, the point value for inserting a boundary in that location was 0.75 and was 0.25 for not inserting a boundary. We calculated the total and average values for relative frequencies of agreement, the latter of which reflects the average proportion of agreement across all potential boundaries. As can be seen in the summary statistics in Table 2, all participants scored quite high with very little variation in agreement rates.

Table 2.

Clausal chunking task and agreement results

Agreement score
(relative frequency)	Mean	SD	Range	Range of participants
Total	2337.36	44.21	120.3−2409.7	2183.96−2391
Average	0.92	0.01	0.05−0.95	0.86−0.94

We also calculated the overall agreement rate of our participants across all extracts using the “irr” package for Fleiss’ kappa (Gamer et al., 2019). This measure ranges from 0 to 1 and assesses interparticipant agreement while also accounting for chance agreement that might arise from random responses or guessing (Fleiss, Levin, & Paik, 2003). As can be seen in Table 3, the value for Fleiss’ kappa was significant, indicating that our participants largely converged on the same location of clausal chunk boundaries.

Table 3.

Agreement and chunk length results

Number of boundaries	Participants	Fleiss’ kappa	p‐value	Average chunk length	Range of chunk lengths	Number of chunks
2530	69	0.54	<.001	5.50 (3.01)	1−19	460

To identify significant chunk boundaries, we conducted a permutation test following Dobrego et al. (2023) and Vetchinnikova, Konina, Williams, Mikušová, and Mauranen (2022), which is available in the ChunkitApp software. This test randomly permuted individual markings (0 for no boundary, 1 for boundary) at each potential boundary one million times (with replacement) and allowed us to estimate the probability of our results being random. The frequency of participant markings at each boundary was compared to the permuted markings using a two‐tailed p‐value. Boundary frequencies that occurred higher than chance were statistically significant, with upper right‐tail values indicating a significant chunk boundary and lower left‐tail values indicating a significant nonchunk boundary. These significant chunk boundaries were used to calculate chunk length and number of chunks. In line with our construal of this task as capturing clausal boundaries, the average chunk length was 5.50 words (SD = 3.01). Note that while we calculated 460 chunks across all extracts, our participants identified 410 clausal chunks, with the remaining 50 chunks reflecting the end of each speech extract. We also note that internal consistency across the 50 extracts (measured using Cronbach's alpha) was relatively low (α = 0.62), although this estimate was not reported in previous studies.

4.3. Regression models

We ran two bivariate correlational models to test if target word or target trigram recall predicted agreement rate on the CC task. We ran separate models with target word and target trigram recall as predictors and CC agreement rate as the response variable in both models. We again note that our use of target recall scores diverges from our preregistered plan to use difference scores (experimental − control) due to low internal consistency of the difference scores. Crucially, as related to our research question, the models showed no significant effect of target word recall (t(68) = 1.63, SE = 0.001, p = .107; see Fig. 2a) or target trigram recall (t(68) = 1.72, SE = 0.003, p = .089; see Fig. 2b) on agreement rate of CC.

Fig. 2.

Scatterplots illustrating the lack of correlation between the multiword recall task and the clausal chunking task: the average relative frequency of agreement as a function of (a) the average target word recall and (b) the average target trigram recall.

5. Discussion

In this study, we aimed to test predictions from the single‐ versus multiple‐mechanisms perspectives on the nature of chunking in language. Specifically, we tested whether chunking performance at the multiword level as measured by the MWC task correlated with performance at the clausal level as measured by the CC task. We observed no correlation between the two tasks, even considering two different measures of multiword recall. Thus, our data suggest that MWC recall does not appear to engage the same chunking processes as CC, as assessed by the CC task. This is consistent with the prediction from the multiple‐mechanism perspective and, more generally, with the chunk‐and‐pass framework (Christiansen & Chater, 2016). Our results also dovetail with similar findings in SL, pointing to a multiple‐mechanism view (see Bogaerts et al., ²⁰²², for a review), and are suggestive of a close relationship between chunking and SL (see also Christiansen, ²⁰¹⁹; Isbilen et al., ²⁰²⁰).

Our findings, however, are subject to several limitations, mostly related to the CC task. Importantly, the average agreement rates on the CC task were at near ceiling, with the lowest rate of agreement at 86% across all extracts resulting in little variation in individual differences for this measure. We note that the lack of individual variation on this task makes it well‐suited for group‐level studies, which is how it was originally used by its developers (Dobrego et al., 2023; Vetchinnikova et al., ^{2017, 2022, 2023}). However, given the minimal variability in scores, it is likely that any meaningful difference in CC among native speakers may be too small to detect with the current measures. We suggest that such high agreement rates may be driven by the top‐down and reflection‐based nature of the CC task, which may not tax memory processes in the way that typical online language processing does (cf. Christiansen, ²⁰¹⁹). Indeed, participants formed quite large clausal chunks (up to 19 words long), supporting the notion that the CC task appears to capture a higher clausal level of representation, compared to the multiword level of the recall task (Christiansen & Chater, 2016), and that these larger chunks may be indicative of more reflection‐based processing with relatively little memory engagement. We thus note that although our results are consistent with a multiple‐mechanism account, further studies are needed to definitively distinguish between the two accounts by comparing linguistic chunking tasks that more directly isolate online language processing.

Moreover, a notable difference in replicating the CC task in Dobrego et al. (2023) is that chunk size was smaller for our participants (M = 5.5) compared to both their native (M = 7.32) and non‐native (M = 7.61) speaking participants. This may be due to potential differences in how chunk size was calculated. Although we followed their scoring method, it may be possible that certain chunk sizes were misrepresented given the flexibility of how language might be chunked. For example, (1a) shows a sample from an extract containing three significant boundary marks (“|”) and thus three resulting chunks.

• (1a) the origin of the wings of the birds | is the same | genetically |

• (1b) the origin of the wings of the birds | is the same genetically |

However, it is possible that most participants intended to form only two chunks as in (1b) given a disproportionate number of boundary inserts. The third boundary mark in (1a) was marked by more than three times as many participants compared to the second boundary, yet all significant boundaries were treated as marking the end of a chunk. Thus, as in (1a), this scoring method resulted in three chunks consisting of eight, three, and one word, respectively; however, most participants formed two chunks of word length eight and four, as illustrated in (1b), which was not captured in the current scoring method. This highlights the need to capture the potential variability of such chunk processing. Future iterations with this task might consider more graded forms of assessing chunk size, perhaps based on the proportion of boundaries inserted at each significant chunk, to gain a more flexible and representative understanding of CC.

6. Conclusion

Our results indicated that MWC recall ability did not correlate with CC, consistent with a multiple‐mechanisms view of chunking and the predictions put forth in the chunk‐and‐pass framework for language processing. We replicated the findings from the CC task in Dobrego et al. (2023), noting limitations in this task as a measure of individual differences in CC. Our finding that agreement rate was near ceiling on the CC task, while performance on the MWC recall task showed a more variable distribution, highlights the possibility that CC might primarily capture more reflection‐based high‐level meaning units and less bottom‐up language processing. Nevertheless, given that the CC task has been used across multiple studies (Anurova et al., 2022; Dobrego et al., ²⁰²³; Vetchinnikova et al., ^{2017, 2022, 2023}), the current findings may help inform future research employing this or similar tasks. More broadly, our study further underscores the methodological observation (e.g., Frinsel & Christiansen, ²⁰²⁴; Hedge, Powell, & Sumner, ²⁰¹⁸) that tasks that are robust at the group level may not necessarily provide reliable measures at the level of individual differences. We hope that our study may prompt the development of new, psychometrically stable individual‐differences measures of chunking that can be used to further differentiate between the single‐ versus multiple‐mechanisms view of language processing.