Quantity Convergence, Quality Divergence: Disentangling Fluency and Accuracy in L2 Mandarin Prosody
| Authors | Yuqi Shi et al. |
| Year | 2026 |
| Field | NLP |
| arXiv | 2602.23071 |
| Download | |
| Categories | cs.CL, cs.AI |
Abstract
While second language (L2) learners may acquire target syntactic word order, mapping this syntax onto appropriate prosodic structures remains a persistent challenge. This study investigates the fossilization and stability of the L2 syntax-prosody interface by comparing 67 native Mandarin speakers with 67 Vietnamese learners using the BLCU-SAIT corpus. By integrating C-ToBI boundary annotation with Dependency Grammar analysis, we examined both the quantity of prosodic boundaries and their mapping to syntactic relations. Results reveal a non-linear acquisition: although high-proficiency learners (VNH) converge to the native baseline in boundary quantity at the Major Phrase level (B3), their structural mapping significantly diverges. Specifically, VNH demote the prosodic boundary at the Subject-Verb (SBV) interface (Major Phrase B3 -> Prosodic Word B1), while erroneously promoting the boundary at the Verb-Object (VOB) interface (Prosodic Word B1 -> Major Phrase B3). This strategy allows learners to maintain high long phrasal output at the expense of structural accuracy. This results in a distorted prosodic hierarchy where the native pattern is inverted.
Engineering Breakdown
Plain English
This paper investigates how second language learners of Mandarin Chinese acquire prosodic structure—the rhythm and intonation patterns of speech—even when they've mastered basic word order. The researchers compared 67 native Mandarin speakers to 67 Vietnamese learners using a corpus of real speech, analyzing both how many prosodic boundaries learners produce and whether those boundaries align with grammatical structure. They found a striking non-linear pattern: advanced Vietnamese learners (VNH) match native speakers in the quantity of major prosodic boundaries, but their placement diverges significantly from native patterns—they systematically demote boundaries at grammatically important positions like subject phrases. This reveals that fluency in quantity doesn't guarantee accuracy in structure, a finding that challenges standard acquisition models.
Core Technical Contribution
The core contribution is empirically demonstrating the syntax-prosody interface as a persistent fossilization point in L2 acquisition, showing that structural mapping (quality) can diverge even when quantitative measures (quantity) converge. The technical novelty combines two annotation frameworks—C-ToBI (Chinese Tonal Obligatory Boundary Indices) for prosody and Dependency Grammar for syntax—to directly measure the mapping relationship between prosodic boundaries and syntactic constituents, rather than treating them separately. This integrated approach reveals that previous studies relying on quantity-only metrics masked systematic divergence in where boundaries attach. The finding contradicts the implicit assumption that high proficiency necessarily means native-like prosodic structure, opening a new diagnostic pathway for detecting fossilization in L2 learners who otherwise sound fluent.
How It Works
The methodology operates in distinct analysis stages: (1) Audio data from native and learner speakers is annotated with C-ToBI labels, which mark prosodic boundaries at multiple hierarchical levels (from lowest B1 to major phrase B3), capturing where the speaker pauses or marks intonational phrases; (2) The same speech is independently parsed with Dependency Grammar, which assigns each word grammatical roles (subject, object, verb, etc.) and syntactic relationships; (3) For each speaker group, the researchers compute two metrics—boundary quantity (raw count of B3-level boundaries per utterance) and boundary accuracy (percentage of B3 boundaries that correctly align with syntactically-predicted positions, particularly at subject phrase edges); (4) Statistical comparison reveals that VNH learners produce B3 boundaries at native-like frequencies but place them in non-native positions, indicating they've internalized the production rate but not the structural mapping rule. The key insight comes from separating quantity from quality: advanced learners hit the right statistical distribution but miss the underlying linguistic constraint.
Production Impact
For engineers building prosody modeling systems—speech synthesis, accent detection, or learner assessment tools—this work directly impacts how you evaluate and train models. A quantitative-only metric (e.g., boundary count, overall F1 score) would classify Vietnamese learners as near-native, but fine-grained structural metrics would flag systematic errors, which is critical for educational feedback systems or quality control in synthetic speech. In a production L2 assessment pipeline, you'd need to implement dual-track evaluation: one measuring overall prosodic patterns (where learners genuinely converge) and another measuring position-specific accuracy relative to syntactic structure (where divergence emerges). This requires building dependency-aware boundary prediction models rather than sequence-level prosody models, adding computational cost (syntax parsing overhead) but preventing the silent failures where learners produce fluent-sounding but structurally inaccurate speech. The trade-off is higher annotation complexity during training data preparation and more sophisticated evaluation code, but the payoff is much better diagnostic power for intermediate-to-advanced learner assessment.
Limitations and When Not to Use This
This study is limited to Vietnamese speakers learning Mandarin—the findings may not generalize to learners with radically different L1 prosody systems (e.g., tone-based languages like Yoruba, or stress-based languages like English), and the paper doesn't explain the mechanism driving the divergence, only documenting it. The BLCU-SAIT corpus contains only 67 speakers per group, which is modest for modern deep learning approaches; larger datasets would be needed to train neural prosody models that capture these divergence patterns. The paper focuses on read speech in controlled settings, not naturalistic conversation where pragmatic and interactive pressures might shift prosodic choices, limiting applicability to real-world deployment. Additionally, the study doesn't provide ablation analysis on the C-ToBI and Dependency Grammar choices—it's unclear whether other annotation schemes would reveal the same divergence or whether the finding is an artifact of these specific frameworks, leaving questions about reproducibility and framework independence.
Research Context
This work builds on decades of second-language acquisition research showing that the syntax-prosody mapping is notoriously difficult for learners, even when individual subsystems (syntax or prosody in isolation) are well-developed. The paper extends prior work by Venditti, Hirschberg, and others on prosodic phrasing in L2 speech by introducing the direct quantitative-versus-qualitative distinction, moving beyond descriptive accounts to measurable divergence metrics. It also engages with fossilization theory in SLA (Second Language Acquisition), specifically challenging the assumption that convergence in aggregate statistics implies native-like competence. The findings open a new research direction: investigating whether this divergence is universal across L2 populations, whether it can be remediated through explicit prosody instruction, and whether neural sequence-to-sequence models trained on syntax-prosody pairs can learn the native mapping rules better than traditional acoustic models.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::