Language and nonlanguage factors in foreign language learning: evidence for the learning condition hypothesis

The study addresses a longstanding question: why do children acquire their native language with ease while adults often struggle with foreign languages? Two dominant hypotheses are evaluated: (1) the Fundamental Difference Hypothesis (FDH), which posits that native and foreign language learning rely on fundamentally different mechanisms and are constrained by critical/sensitive periods; and (2) the Linguistic Coding Differences/Deficit Hypothesis (LCDH), which posits shared core linguistic functions across languages leading to correlated outcomes. The authors propose an alternative, the Learning Condition Hypothesis (LCH), which argues that similarities in learning conditions (naturalistic/implicit vs. instructed/explicit) drive associations in proficiency across languages. In Hong Kong, participants typically acquire Chinese (L1) naturalistically, learn English (L2) in school with some ambient input, and learn L3 (French/German/Spanish) in classroom settings. This continuum allows testing whether proficiency associations align with similarity in learning conditions rather than typological proximity or putative fundamental differences. The study aims to test FDH, LCDH, and LCH using a large sample with comprehensive proficiency measures while controlling for nonlanguage factors.

Prior research shows a persistent gap in ultimate attainment and processing efficiency between L1 and L2, interpreted by FDH as reflecting different underlying mechanisms and maturational constraints. Other studies report within-individual correlations between L1 and L2 skills (e.g., decoding, spelling), supporting LCDH’s shared-core-functions view. However, many correlations were shown in typologically similar languages and on metalinguistic tasks. The present work highlights that prior comparisons often overlook learning conditions. The authors also situate their work relative to L3 theories (e.g., typological primacy model, cumulative enhancement model, L2 status factor, dynamic model of multilingualism, revised hierarchical model, linguistic proximity model), noting that these largely focus on morphosyntax and transfer rather than overall proficiency, and that typology may not fully account for proficiency associations when learning contexts differ.

Design and participants: Cross-sectional study of 636 Cantonese (Chinese) L1 speakers (ages ~18–25) at The Chinese University of Hong Kong enrolled in Modern Languages courses, learning French (n=187), German (n=176), or Spanish (n=273) as L3. All learned English as L2 from early childhood primarily in school. Inclusion: normal hearing screening; nonverbal IQ ≥85 (TONI-4). Collected demographics (gender, ages at HKDSE and at L3 testing), family SES (Hollingshead index), musical training (≥1 year vs. <1 year/none), and detailed motivational/affective measures. Measures: - L1 and L2 proficiency: Composite grades from Hong Kong Diploma of Secondary Education (HKDSE) Chinese (L1) and English (L2) subjects (covering reading, writing, speaking, listening). HKDSE is standards-referenced and annually calibrated. - L3 proficiency (Global score): Derived via principal component analysis (PCA) combining laboratory and classroom measures: 1) Narrative production (storytelling “Frog, Where Are You?”) transcribed and analyzed with CLAN/CHAT; 15 indices summarized by PCA (KMO=0.78; 2–3 components explaining ~65%). 2) Language access: picture naming of 31 body parts; accuracy judged by native speakers. 3) Pronunciation ratings: native speakers rated two 20–30 s excerpts on a 9-point native-likeness scale via crowdsourcing; ratings averaged. 4) Classroom exam: end-of-term L3 exam z-scores within class. All L3 measures (standardized within class level and language) plus exam scores were combined via PCA (varimax; KMO=0.60; Bartlett p<0.001). The first component (eigenvalue=1.58; 32% variance) was taken as the L3 Global score. - Motivation and affect: Modern Language (ML) learner questionnaire adapted and reduced via PCA. Part I (49 items): external motivation (e.g., ought-to self, family influence) and internal motivation (ideal L2 self) components (KMO=0.93). Part II (17 items): attitude (language, culture, community) and anxiety components (KMO=0.86). Context/learning conditions: L1 acquired naturalistically; L2 learned in school with some ambient exposure; L3 learned in classroom with instruction in L3/English; teachers were native or near-native L3 speakers. This establishes an implicit-to-explicit continuum from L1→L2→L3. Analytic strategy: - Bivariate associations: Spearman correlations among L1, L2, L3 Global; FDR correction; pairwise deletion for missingness. - General linear models (GLM): • Model 1 DV=L2 (English HKDSE). Predictors: L1 (Chinese HKDSE), gender, musical training, family SES, nonverbal IQ, HKDSE age. FDR-adjusted p-values. • Model 2 DV=L3 Global. Predictors: L1 (Chinese), L2 (English), gender, musical training, family SES, nonverbal IQ, L3 age, internal/external motivation, attitude, anxiety. FDR-adjusted p-values. - Structural equation modeling (SEM): Latent variables for L1, L2, L3 proficiencies indicated by the respective exam/global scores; included nonlanguage covariates. Two models tested: LCH paths (L1→L2, L2→L3) and LCDH model adding L1→L3. Robust estimation with FIML for missing data; fit indices (RMSEA, SRMR, CFI, TLI) reported; model comparison via Δχ². Subgroup SEMs by L3 class level (low vs high) and by L3 language (French, German, Spanish) to probe typological effects. - Machine learning (SVR): Linear SVR (C=1, epsilon=0.1) predicting L3 Global and L2 HKDSE. Nested 10-fold cross-validation with 10,000 iterations; performance quantified by Pearson correlation between predicted and observed outcomes; permutation-based null distributions. Predictor sets: all predictors; only L2 for L3 prediction; only L1 for L3 prediction; only L1 for L2 prediction. Power: Preliminary data from first 25 learners per L3 suggested r≈0.25 between L1 and L2. With family-wise α=0.05 (Bonferroni p=0.017 for three tests), minimum N≈163 was required; key measures had Ns ≥167 per language group, ensuring adequate power.

- Bivariate correlations (FDR-corrected): L1 (Chinese) vs L2 (English) r=0.26, p<0.001; L2 vs L3 Global r=0.28, p<0.001; L1 vs L3 Global r=0.05, p=0.263 (ns). This pattern aligns with LCH. - GLM predicting L2 (English HKDSE): Significant predictors included L1 (Chinese) (Estimate≈0.24; ΔR²=0.06; p<0.001), musical training (β=0.31, p=0.015), family SES (β=0.02, p<0.001), and HKDSE age (β=−0.16, p=0.015). Gender and nonverbal IQ were not significant. Model R²=0.165 (adjusted R²=0.154), p<0.001. - GLM predicting L3 Global: L2 (English) was the strongest predictor (Estimate≈0.28; ΔR²=0.06; p<0.001). Attitude (β=0.13, p=0.043) and L3 age (β=−0.08, p=0.044) were significant; L1 (Chinese) was not (β=−0.02, p=0.580). Model R²=0.166 (adjusted R²=0.148), p<0.001. - SEM (all participants): Both LCH and LCDH models fit acceptably (LCH: RMSEA=0.025 [0.000–0.054], SRMR=0.021, CFI=0.971, TLI=0.946; LCDH: RMSEA=0.028 [0.000–0.058], SRMR=0.021, CFI=0.967, TLI=0.932). The L1→L3 path was not significant (b=−0.021, 95% CI −0.101 to 0.059). Model comparison showed no improvement by adding L1→L3 (Δχ²=0.269, p=0.604), favoring the parsimonious LCH. Significant paths: L1→L2 (b≈0.257, p<0.001), L2→L3 (b≈0.210–0.217, p<0.001). Covariate effects included SES→L2 (positive), HKDSE age→L2 (negative), attitude→L3 (positive), anxiety→L3 (negative), etc. (see Table 4 in text). - Subgroup SEMs: No significant differences between low vs high L3 proficiency groups (Δχ²=11.62, p=0.637). No significant differences across L3 language groups (German vs French: Δχ²=22.30, p=0.073; German vs Spanish: Δχ²=16.73, p=0.271; Spanish vs French: Δχ²=11.73, p=0.628), suggesting typological distance did not modulate the L2→L3 association in this dataset. - SVR prediction: • L3 Global: With all predictors, predictability mean cc=0.355 (SD=0.039) vs null mean≈0.001 (SD=0.068), p<0.001, Cohen’s d=6.39. With only L2, mean cc=0.278 (SD=0.039), p<0.001, d=5.22. With only L1, mean cc=0.052 (SD=0.042), p<0.001, d=0.96 (much smaller effect). • L2 (English): With all predictors, mean cc=0.373 (SD=0.036), p<0.001, d=6.45. With only L1 (Chinese), mean cc=0.257 (SD=0.037), p<0.001, d=7.17. Overall, results converge to significant associations between L1–L2 and L2–L3, but not L1–L3, consistent with the Learning Condition Hypothesis.

Findings show that proficiency correlations align with similarity in learning conditions rather than simple nativeness or typological proximity. L1 (naturalistic) is associated with L2 (mixed/partially implicit), and L2 (classroom-based with some ambient exposure) is associated with L3 (classroom-based), whereas L1 is not associated with L3. This pattern supports the Learning Condition Hypothesis over FDH (which predicts no L1 links to L2/L3) and over LCDH (which predicts universal correlations among L1, L2, L3). The associations persisted after controlling for nonlanguage factors (SES, musical training, age, gender, nonverbal IQ, motivation, attitude, anxiety) and were corroborated by GLM, SEM, and cross-validated SVR, enhancing robustness and generalizability. Subgroup analyses indicated the pattern was stable across proficiency levels and across L3 languages (French, German, Spanish), and not detectably modulated by typological distance in this context. The results are compatible with neurocognitive accounts positing reuse of learning mechanisms aligned with exposure and instruction regimes, and are not easily explained by critical period effects alone (given early L2 onset and the stronger L2–L3 link). The work reframes native vs foreign language learning differences as largely reflecting differences in learning conditions, with implications for optimizing language teaching and understanding variability in multilingual attainment.

This large-scale study introduces and supports the Learning Condition Hypothesis: similarities in learning conditions across languages predict associations in proficiency. Using comprehensive measures and multiple analytic approaches, the authors show significant L1–L2 and L2–L3 associations but no L1–L3 association among Cantonese L1 learners of English (L2) and French/German/Spanish (L3). The results suggest that learning conditions may be a principal factor shaping cross-language proficiency relations, beyond typological proximity or putative fundamental differences. Future research should: (1) incorporate broader cognitive and aptitude measures (procedural/declarative memory, working memory, language aptitude); (2) examine populations where native and nonnative languages share learning contexts (e.g., immigrants, heritage speakers) to further test LCH; (3) disentangle age of learning from learning conditions more directly; and (4) consider potential genetic contributions to language learning variability.

- Unmeasured cognitive constructs (procedural/declarative memory, working memory, language aptitude) were not directly assessed and may contribute to L3 outcomes. - Age of learning covaries with learning conditions (L1 from birth, L2 ~3 years, L3 ~18 years), potentially confounding effects despite analyses suggesting age was not the primary driver. - Typological distance was approximated by ancestral relationships; psychotypology was not directly measured, and no standardized L3 proficiency exam was available (L3 Global was derived from multiple measures via PCA). - Genetic variation potentially influencing language learning was not examined. - Generalizability is to similar educational contexts; findings in immersion/heritage contexts should be tested.