How classifiers facilitate predictive processing in L1 and L2 Chinese: the role of semantic and grammatical cues

ABSTRACT

This study examines long-standing claims that L2 learners rely more on non-grammatical than on grammatical information during sentence processing compared to native speakers. Nominal classifiers in Mandarin Chinese offer an ideal opportunity to test this claim, as they simultaneously encode semantic as well as grammatical form-class cues about co-occurring nouns. This paper reports findings from a visual world eye-tracking experiment with L1 and L2 speakers of Mandarin, which was designed to assess listeners’ relative reliance on these two concurrently available cues when creating expectations about an upcoming noun in a sentence. Results show that L2 listeners experienced greater competition than L1 listeners from nouns that were grammatically incompatible with the classifier they heard but shared semantic features associated with it. The greater reliance on semantic cues observed in L2 processing is argued to be an effect of adaptation to the relative reliability of information, serving to maximise L2 processing efficiency.

KEYWORDS:

• Classifier
• Mandarin Chinese
• eye-tracking
• cue weighting
• prediction
• L2 processing

Acknowledgments

We thank Craig Chambers for sharing materials from Tsang and Chambers (2011), and Victoria Lee for assistance with data collection. We are also grateful for valuable feedback from the audiences at CUNY 2017, BUCLD 2017, and ISBPAC 2018, as well as from three anonymous LCN reviewers.

Disclosure statement

No potential conflict of interest was reported by the authors.

ORCID

Theres Grüter http://orcid.org/0000-0001-6354-9787

Notes

1 A preliminary short report of this study appeared in the BUCLD proceedings (Grüter, Lau, & Ling, 2018) while data collection was still on-going. The small sample of L2 participants at that point, which also included some heritage speakers (now excluded), did not afford the statistical power needed for the planned analyses (reported here, Analysis 2), nor did this format allow for presentation of materials, including norming and rating procedures, sufficient for replicability, or full discussion of theoretical implications.

2 The in-house listening comprehension task consisted of an adaptation of the listening part of the Hanyu Shuiping Kaoshi (HSK or Chinese Standard Exam) Level Two sample test available on the Confucius Institute website (http://www.confuciusinstitute.manchester.ac.uk/hsk/hsk-learning-resources/past-papers-hsk-2/).

3 Note that unlike in T&C’s Cantonese stimuli, the classifier in our Mandarin stimulus sentences did not immediately precede the noun, but preceded both the copula shì and the noun. This word order was selected to increase the critical time window between classifier and noun onset, considering the possibility that potential effects of prediction might arise more slowly in L2 listeners (Kaan, 2014).

4 T&C’s experiment included visual scenes with four objects: a target, a competitor, and two distractors. In order to maximise potential looks to competitors, visual scenes in the present study included only one distractor, i.e., three objects in total. If the Mandarin translation of both Cantonese distractor items fit our criteria, Distractor 1 was selected.

5 Two trials with no recorded fixation data during the critical time window were removed.

6 For example, if on a given trial, of the 58 bins comprised by this window, 20 contained looks to the target and 15 contained looks to the competitor, the TargetAdvantage score for this trial would equal 5. This dependent measure was chosen over weighted empirical logits (Barr, 2008) because it was more normally distributed.

7 This led to the removal of 37 trials (L1: 6/283, L2: 31/443) in which an early initiated fixation continued throughout the critical time window.

8 All models reported contain random intercepts only. Models with random slopes were attempted, but did not consistently converge. For the sake of comparison, we therefore modelled all data with random intercepts only. When fuller models converged, anova() comparisons indicated that they were not a significantly better fit for the data. The patterns of results did not change in fuller models.

Additional information

Funding

This research was supported by funding from Language Learning’s Small Grant Research Program.