L2 Processing Advantages of Multiword Sequences: Evidence from Eye-Tracking

Elma Kerz, Arndt Heilmann, Stella NeumannACL 2019

TL;DR

This study investigates whether non-native English speakers (L2 learners) develop the same sensitivity to the frequency of multiword sequences (MWS) as native speakers. Using eye-tracking, the research found that L2 learners process common MWS faster, suggesting that similar underlying mechanisms drive language processing in both native and non-native speakers, supporting emergentist views of language acquisition.

What Problem Does This Paper Address?

For decades, language processing models have largely operated under a "words and rules" paradigm, where language is understood as a combination of individual words governed by grammatical rules. However, a growing body of research suggests that language users are highly attuned to the statistical regularities of language, including the frequency of word sequences. This challenges the strict separation between lexicon and grammar.

While native speakers demonstrably benefit from the frequency of multiword sequences (MWS) – common phrases like "in order to" or "a lot of" – it's less clear if intermediate-to-advanced L2 learners develop this same sensitivity. Previous studies on L2 learners have yielded mixed results, and often relied on less ecologically valid methods like self-paced reading or acceptability judgments. This paper aims to replicate the findings of MWS frequency advantages in native speakers using a more naturalistic eye-tracking study with L2 English learners, specifically focusing on compositional MWS (those that are syntactically regular and semantically transparent, not just idiomatic).

Methods

The study employed an eye-tracking experiment to measure how L2 English speakers process sentences containing common multiword sequences (lexical bundles, LBs) compared to sentences with less common, matched sequences (non-lexical bundles, NLBs).

Key eye-movement measures were recorded:

  • First Fixation Duration (FFD): Time spent on the initial fixation within the MWS region.
  • First Pass Reading Time (FPRT): Total time spent fixating within the MWS region before moving past it.
  • Total Reading Time (TRT): Sum of all fixations within the MWS region, including regressions.
  • COUNT: Number of regressive saccades into the MWS region.

Statistical analysis involved fitting mixed-effects regression models to these measures. The primary predictor was "lexical bundlehood" (LB vs. NLB). Control variables included the length of the MWS and the participants' L2 proficiency (measured by LexTALE scores). Fixation times were log-transformed to normalize distributions.

Datasets

The study utilized stimulus material previously developed by Tremblay et al. (2011). This material consisted of pairs of short sentences (average 8.5 words) that differed by only one word. The critical difference was the presence of a multiword sequence that met Biber and Conrad's (1999) criteria for a "lexical bundle" (LB) versus a matched non-lexical bundle (NLB). The frequency of these MWS was determined from the spoken subcorpus of the British National Corpus (BNC). The stimulus set included 40 sentences containing LBs, 40 containing NLBs, and 40 filler sentences (20 meaningful, 20 nonsensical).

Key Insights and Ideas

  • L2 learners are sensitive to MWS frequency: The study provides strong evidence that intermediate-to-advanced L2 English speakers process common multiword sequences significantly faster than less common ones. This challenges the notion that such statistical learning is exclusive to native language acquisition.
  • Unified processing mechanisms: The findings support emergentist theories of language, which propose that language acquisition and processing rely on general cognitive mechanisms rather than innate, language-specific modules. The observed similarities in processing advantages between L1 and L2 speakers suggest a unified system.
  • Chunking as a key mechanism: The sensitivity to MWS frequency is linked to "chunking" – the process of grouping frequently co-occurring words into larger units. This chunking helps overcome the limitations of human memory and processing speed, allowing for more efficient language comprehension.
  • Implications for L1/L2 learning: The results suggest that the differences between L1 and L2 learning might be less pronounced than previously thought, particularly for advanced learners. This points towards the need for unified theoretical models that can account for both.

Key Results

Mixed-effects modeling revealed significant effects for lexical bundlehood across most measures:

  • Total Reading Time (TRT): Participants were significantly faster processing sentences with LBs. On average, sentences with LBs were read 339 ms faster than those with NLBs.
  • First Pass Reading Time (FPRT): LBs led to 181 ms faster processing.
  • COUNT (Regressive Saccades): Participants made significantly fewer regressive saccades into LB regions (an average increase of 1.52 regressive saccades for NLBs compared to LBs).
  • First Fixation Duration (FFD): The effect for FFD was only marginally significant (p < 0.07), suggesting that while overall processing is faster, the initial engagement with the sequence might not be as strongly affected.

These results are consistent with previous findings in native speakers (e.g., Tremblay et al., 2011), who reported a 380ms advantage for LBs in sentence-by-sentence reading.

Limitations and Future Work

The authors acknowledge several avenues for future research:

  • Exploring more complex statistics: The current study focused on simple corpus-derived frequencies. Future work could investigate sensitivity to more complex distributional statistics like transitional probability, mutual information, or entropy.
  • Written vs. Spoken Input: The stimulus material was derived from spoken language corpora. Given the growing evidence for the role of written input, it would be valuable to explore how users tune to statistics in different registers and genres of written text.
  • Individual Differences: The study did not deeply explore individual differences in sensitivity to MWS. Future research could investigate how factors like working memory, cognitive abilities, and language learning experiences influence this sensitivity.

Summary updated

BibTeX
@inproceedings{kerz2019l,
  author = {Elma Kerz and Arndt Heilmann and Stella Neumann},
  title = {{L2 Processing Advantages of Multiword Sequences: Evidence from Eye-Tracking}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2019},
}
Figure