Neither Here Nor There: Cross-Lingual Representation Dynamics of Code-Mixed Text in Multilingual Encoders

Debajyoti Mazumder; Divyansh Pathak; Prashant Kodali; Jasabanta Patro

cs.CLMar 20, 2026

Neither Here Nor There: Cross-Lingual Representation Dynamics of Code-Mixed Text in Multilingual Encoders

Authors: Debajyoti Mazumder, Divyansh Pathak, Prashant Kodali, Jasabanta Patro

Organizations: Indian Institute of Science Education and Research, Bhopal, India · Microsoft Corporation

Abstract

Multilingual encoder-based language models are widely used for code-mixed analysis, yet their internal representations of code-mixed inputs -- and their relationship to the constituent languages -- remain poorly understood. Using Hindi-English as a case study, we construct a unified trilingual corpus of parallel English, Hindi (Devanagari), and Romanized code-mixed sentences. We then probe cross-lingual representation alignment in standard multilingual encoders and their code-mix-adapted variants using CKA, token-level saliency, and entropy-based uncertainty analysis. We find that while standard models align English and Hindi well, code-mixed inputs remain loosely connected to either language -- and that continued pre-training on code-mixed data improves English-code-mixed alignment at the cost of English-Hindi alignment. Interpretability analyses further reveal a clear asymmetry: models process code-mixed text through an English-dominant semantic subspace, while native-script Hindi provides complementary signals that reduce representational uncertainty. Motivated by these findings, we introduce a trilingual post-training alignment objective that brings code-mixed representations closer to both constituent languages simultaneously, yielding more balanced cross-lingual alignment and downstream gains on sentiment analysis and hate speech detection -- showing that grounding code-mixed representations in their constituent languages meaningfully helps cross-lingual understanding. Code is available at: https://github.com/debajyotimaz/tri_align_EMNLP_2026.

Figures & tables

Figure 1: Vanilla multilingual model versus code-mixed adapted multilingual model. Here, S represents similarity.

Dataset	# Samples	Languages
CM-En parallel Dhar et al. (2018)	6,096	CM, EN
PHINC Srivastava and Singh (2020)	13,738	CM, EN
LINCE 2021 1	8060	CM, EN, HI

Table 1: Summary of the datasets used in this work, where EN = English, HI = Hindi (Devanagari script), and CM = Hindi (Roman-script) - English code-mixed text.

Model	Version
mBERT	bert-base-multilingual-cased
Hing-mBERT	l3cube-pune/hing-bert
Hing-mBERT-Mixed	l3cube-pune/hing-bert-mixed
XLM-R	xlm-roberta-base
Hing-RoBERTa	l3cube-pune/hing-roberta
Hing-RoBERTa-Mixed	l3cube-pune/hing-roberta-mixed

Table 2: Multilingual encoder models used in our experiments. Hing models denote mBERT and XLM-R variants further trained on the L3Cube-HingCorpus.

Figure 2: Directional cross-lingual alignment accuracy across model families. Solid bars show retrieval from language A $\rightarrow$ B, hatched bars show B $\rightarrow$ A.

Figure 3: t-SNE visualization of sentence representations using mBERT and XLM-R-family models.

Figure 4: Directional dot-product retrieval accuracy across encoder models for EN–HI, EN–CM, and HI–CM, averaged across all layers. Solid bars denote A $\rightarrow$ B and hatched bars denote B $\rightarrow$ A.

Figure 5: Layer-wise CKA alignment for mBERT and XLM-R family. Each subplot shows cross-lingual representation alignment for EN-CM , EN-HI , and HI-CM across layers.

Figure 6: Language-wise per-token saliency for encoder-only models on Hinglish inputs. For each model, we report the average Relative Importance (RI) score assigned to English and Hindi tokens Special tokens (e.g., <bos> , <eos> ) are excluded.

Figure 7: Uncertainty reduction in code-mixed (CM) representations across encoder layers in MLMs.

Last Layer Retrieval
	EN $\leftrightarrow$ CM		EN $\leftrightarrow$ HI		HI $\leftrightarrow$ CM		CLAS
Model	$\rightarrow$	$\leftarrow$	$\rightarrow$	$\leftarrow$	$\rightarrow$	$\leftarrow$
mBERT Family
mBERT	54.75	42.90	74.87	33.80	26.30	39.05	14.20
mBERT Trilingual	69.43	63.49	71.52	73.23	54.49	49.69	50.97
Hing-mBERT	52.78	71.86	32.30	33.43	24.70	28.99	17.01

Table 3: Bidirectional cross-lingual alignment accuracy (%) using dot-product similarity with 10 negative samples evaluated on the test split. All scores are average of three random seeds.

Model	Sentiment	Hate
mBERT	0.4012	0.5220
mBERT-Tri	0.4898	0.5953
XLM-R	0.5720	0.6307
XLM-R-Tri	0.5986	0.6701

Table 4: Mean consistency scores across CM, EN, and HI training settings for sentiment analysis and hate speech detection. The definition of consistency metric and additional scores are presented in Appendix F .

Appendix figures & tables18 assets

Supplementary material from the paper’s appendix.

Appendix

Metric	Score
Average BLEU	0.2527
Average SacreBLEU	38.73
Average COMET	0.8347

Appendix

Table 5: Backtranslation-based evaluation of the Hindi translations used in the trilingual corpus.

Last layer
	EN-CM	CM-EN	EN-HI	HI-EN	HI-CM	CM-HI	CLAS
mBERT (Google Translate)	54.75	42.90	74.87	33.80	26.30	39.05	14.20
XLM-R (Google Translate)	50.56	50.67	58.81	54.33	38.37	40.46	39.53
mBERT (IndicTrans2)	55.37	42.98	72.76	32.46	25.35	37.79	13.53
XLM-R (IndicTrans2)	50.15	50.55	58.34	54.01	38.17	40.44	39.28
Mean layer

Appendix

Table 6: Comparison of cross-lingual alignment scores obtained using Hindi translations generated by Google Translate and IndicTrans2.

Model	Version
IndicBERT-v2	ai4bharat/IndicBERTv2-MLM-only
Qwen3-Embedding	Qwen/Qwen3-Embedding-0.6B

Appendix

Table 7: Additional multilingual encoder models used in our experiments.

Model	Last Layer	Mean
IndicBERTv2	33.59	34.14
IndicBERTv2 Tri	51.47	37.27
Qwen3-Embedding	22.08	50.25
Qwen3-Embedding Tri	79.59	67.70

Appendix

Table 8: Dot-product alignment scores for IndicBERT-v2, its trilingual variant, Qwen3-Embedding-0.6B, and its trilingual variant. Reported values correspond to CLAS.

Hate Speech Detection
Task	Model	Train CM	Train EN	Train HI
	IndicBERTv2	0.6340	0.6097	0.6142
	IndicBERTv2 Tri	0.6576	0.6438	0.6496
Sentiment Analysis
	IndicBERTv2	0.6346	0.4932	0.6063
	IndicBERTv2 Tri	0.6393	0.5014	0.6154

Appendix

Table 9: Cross-lingual consistency scores for IndicBERT-v2 on hate speech detection and sentiment analysis.

CLAS
Model Family	Training Objective	Last Layer	Mean
mBERT Family
mBERT	Standard	14.20	66.36
mBERT	Tri	50.97	71.02
mBERT	InfoNCE	86.78	70.07
mBERT	AnInfoNCE	83.79	73.20

Appendix

Table 10: Comparison of alignment objectives using dot-product retrieval. Reported values correspond to CLAS.

Model	Train CM	Train EN	Train HI
mBERT	0.5542	0.4805	0.5314
mBERT Tri	0.6194	0.6119	0.5545
mBERT InfoNCE	0.6207	0.5627	0.5903
mBERT AnInfoNCE	0.5971	0.5700	0.6100
XLM-R	0.6217	0.6326	0.6379
XLM-R-Tri	0.6794	0.6732	0.6577

Appendix

Table 11: Cross-lingual consistency scores for hate speech detection under alternative alignment objectives.

Model	Train CM	Train EN	Train HI
mBERT	0.3283	0.3674	0.5079
mBERT Tri	0.4464	0.4710	0.5519
mBERT InfoNCE	0.5433	0.4657	0.5760
mBERT AnInfoNCE	0.5553	0.4562	0.5550
XLM-R	0.6304	0.4686	0.6170
XLM-R-Tri	0.6618	0.5022	0.6318

Appendix

Table 12: Cross-lingual consistency scores for sentiment analysis under alternative alignment objectives.

	EN $\leftrightarrow$ CM		EN $\leftrightarrow$ HI		HI $\leftrightarrow$ CM		CLAS
Model	$\rightarrow$	$\leftarrow$	$\rightarrow$	$\leftarrow$	$\rightarrow$	$\leftarrow$
Last Layer Retrieval
mBERT Family
mBERT	54.75	42.90	74.87	33.80	26.30	39.05	14.20
mBERT Trilingual	69.43	63.49	71.52	73.23	54.49	49.69	50.97
–w/o Alignment	52.90	47.49	66.83	30.40	22.91	36.55	15.06

Appendix

Table 13: Bidirectional alignment scores using dot product similarity. Last layer accuracy (%), mean accuracy (%) and CLAS (%) are reported for model families evaluated on the test split. All scores are average of three random seeds.

Last Layer Retrieval
	EN $\leftrightarrow$ CM		EN $\leftrightarrow$ HI		HI $\leftrightarrow$ CM		CLAS
Model	$\rightarrow$	$\leftarrow$	$\rightarrow$	$\leftarrow$	$\rightarrow$	$\leftarrow$
mBERT Family
mBERT	36.46	28.11	52.07	16.41	11.66	19.06	1.68
mBERT Trilingual	51.73	46.61	57.66	60.80	36.02	31.57	32.70
Hing-mBERT	35.11	46.36	26.85	14.70	12.03	10.72	3.82

Appendix

Table 14: Bidirectional cross-lingual alignment accuracy (%) using dot-product similarity with 100 negative samples evaluated on the test split. All scores are average of three random seeds.

Model Family	Model	PPL (CM)
mBERT	mBERT	462.41
	mBERT Trilingual	119.28
	Hing-mBERT	10.98
	Hing-mBERT Trilingual	8.08
	Hing-mBERT-Mixed	10.40
	Hing-mBERT-Mixed Trilingual	8.37

Appendix

Table 15: Perplexity scores computed on the code-mixed test set.

Label	Train	Val	Test	Total
Sentiment Patwa et al. (2020)
Neutral	5,426	1,124	1,071	7,621
Positive	4,853	979	979	6,811
Negative	4,232	887	876	5,995
Total	14,511	2,990	2,926	20,427
Hate Speech Bohra et al. (2018)

Appendix

Table 16: Dataset statistics for the sentiment analysis and hate speech detection tasks after preprocessing.

Figure 8: Layer-wise CKA alignment for encoders and their trilingual versions. Each subplot shows cross-lingual representation alignment for EN-CM , EN-HI , and HI-CM across layers.

Figure 9: Layer-wise SVCCA alignment for mBERT and XLM-R with their code-mixed adapted models. Each subplot shows cross-lingual representation alignment for EN-CM , EN-HI , and HI-CM across layers.

Train	Model	Test Set			Consistency
Train	Model	CM	EN	HI	Consistency
CM	mBERT	0.5192	0.5672	0.3095	0.3283
	mBERT Trilingual	0.6290	0.6420	0.4281	0.4464
	–w/o Alignment	0.6537	0.6177	0.3538	0.3780
EN	mBERT	0.4196	0.7323	0.4604	0.3674
	mBERT Trilingual	0.5333	0.7277	0.5103	0.4710

Appendix

Table 17: Performance of models on sentiment classification task when trained on different source languages and evaluated on code-mixed (CM), English (EN), and Hindi (HI) test sets. Notation: Except consistency, all others are macro-F1 scores averaged across three random seeds.

Train	Model	Test Set			Consistency
Train	Model	CM	EN	HI	Consistency
CM	Hing-mBERT	0.6916	0.7141	0.3221	0.3558
	Hing-mBERT Trilingual	0.6722	0.6895	0.3784	0.4052
	Hing-mBERT Mixed	0.7241	0.7122	0.6812	0.6837
	Hing-mBERT Mixed Trilingual	0.7139	0.7110	0.6900	0.6919
EN	Hing-mBERT	0.5613	0.7311	0.2748	0.2918

Appendix

Table 18: Performance of Hing-mBERT and Hing-RoBERTa models on sentiment classification task when trained on different source languages and evaluated on code-mixed (CM), English (EN), and Hindi (HI) test sets. Notation: Except consistency, all others are macro-F1 scores averaged across three random seeds.

Train	Model	Test Set			Consistency
Train	Model	CM	EN	HI	Consistency
CM	mBERT	0.6706	0.6194	0.5373	0.5542
	mBERT Trilingual	0.6631	0.6188	0.6427	0.6194
	–w/o Alignment	0.6847	0.6575	0.5937	0.6072
EN	mBERT	0.5777	0.6294	0.4732	0.4805
	mBERT Trilingual	0.6391	0.6362	0.6094	0.6119

Appendix

Table 19: Performance of models on hate speech classification task when trained on different source languages and evaluated on code-mixed (CM), English (EN), and Hindi (HI) test sets. Notation: Except consistency, all others are macro-F1 scores averaged across three random seeds.

Train	Model	Test Set			Consistency
Train	Model	CM	EN	HI	Consistency
CM	Hing-mBERT	0.6981	0.6748	0.3890	0.4468
	Hing-mBERT Trilingual	0.6556	0.6212	0.5136	0.5363
	Hing-mBERT Mixed	0.6961	0.6660	0.6109	0.6224
	Hing-mBERT Mixed Trilingual	0.6915	0.6692	0.6475	0.6514
EN	Hing-mBERT	0.6970	0.6883	0.3890	0.4482

Appendix

Table 20: Performance of Hing-mBERT and Hing-RoBERTa models on hate speech classification task when trained on different source languages and evaluated on code-mixed (CM), English (EN), and Hindi (HI) test sets. Notation: Except consistency, all others are macro-F1 scores averaged across three random seeds.

Explore similar work

Sep 11, 2026cs.CL

IndicTriMix: Developing Language Identification Datasets and Models for Tri-Language Code-Mixing

Language identification in code-mixed text, largely observed in social media, is highly essential when users frequently switch between multiple languages within a single utterance. Accurately identifying the languages of code-mixed tokens becomes an urgent necessity. Traditional language identification models, designed for monolingual text, are not well suited for token-level language identification in code-mixed settings. We formulate the task as a sequence labeling problem and fine-tune contextual transformer-based models MuRIL and XLM-RoBERTa best suited for Indian languages. We evaluate these systems on three different data configurations (Hindi, Gujarati, and Bengali) to predict language labels for individual tokens. We release a benchmark for language identification in code-mixed tokens with manually annotated test sets. We propose two approaches of code-mixed generation using parallel sentences of three languages. The trained models demonstrate the effectiveness of contextual embeddings for token-level language identification in multilingual social media text. For reproducibility and to facilitate future research, we publicly release our fine-tuned models.

Pruthwik Mishra, Rudra Trivedi, Avi Patel +2

Sardar Vallabhbhai National Institute of Technology, Surat, India · TCS Research, Hyderabad, India

Code-Switching Multilingual Language Pairs Motion-Language Alignment Mix Token-Level Uncertainty+1

Ask PDF

Jul 28, 2026cs.CL

Evaluation of forced alignment of code-mixed speech: the case of Hindi-English

Code-mixed speech poses unique challenges to forced alignment: expanded inventories, orthographic errors, and speaker variation. We evaluate forced alignment of Hindi-English code-mixed speech using the Montreal Forced Aligner. We address 2 problems: (1) free variation involving native vs non-native pairs and (2) phonemic boundary detection for mid-utterance English words. Bootstrapping strategies substantially outperform unmodified lexicons. Acoustic models trained on sentence-level code-mixed data achieve a mean error of 4.15ms, ie. ten times lower than monolingual Hindi (38.18ms) or isolated English (37.58ms) alternatives. Principled lexicon design and code-mixed training data are both essential for reliable alignment of bilingual speech.

Ayushi Pandey, Pamir Gogoi, Kevin Tang

Karya, India · Department of English Language and Linguistics, Institute of English and American Studies, Faculty of Arts and Humanities, Heinrich Heine University Düsseldorf, Germany · Department of Linguistics, University of Florida, United States of America

Grapheme-To-Phoneme Indian Languages

Ask PDF

May 28, 2026cs.CL

Evaluating Cross-lingual Knowledge Consistency in Code-Mixed vis-a-vis Indian Languages using IndicKLAR

Large language models often exhibit a substantial gap between their performance in English and in lower-resourced languages on equivalent knowledge queries---a cross-lingual consistency issue that remains underexplored for Indian languages and their code-mixed counterparts. To study this gap, we introduce IndicKLAR, an Indic extension of the KLAR-CLC benchmark covering 18 of the 22 scheduled Indian languages. For 11 widely used language pairs, we additionally provide code-mixed variants. Both monolingual and code-mixed inputs verified by native speakers. This three-way alignment enables us to examine how knowledge recall consistency varies across English, code-mixed, and native Indian language inputs. Across nine open-weight models, we find that the accuracy gap between native-language and English inputs can reach

\sim

0.50, while code-mixed inputs substantially reduce this gap, bringing performance within

\sim

0.05 of English without any model-level intervention. Motivated by this finding, we evaluate several prompting strategies that differ in how explicitly language conversion is exposed: a two-stage translate-then-answer setup, a one-stage joint translation-and-answer prompt, and Translate-in-Thought (TinT)---a single-step strategy in which the model internally converts the input and outputs only the final answer. Across the native

\rightarrow

code-mixed

\rightarrow

English performance trajectory, we observe a consistent flip point---the transition from incorrect to correct prediction---between the native and code-mixed settings. Notably, this pattern holds both when the code-mixed representation is explicitly provided as input or when the model is prompted to convert internally using TinT.

Debajyoti Mazumder, Divyansh Pathak, Prashant Kodali +3

Indian Institute of Science Education and Research, Bhopal, India · Microsoft Corporation · UNSW Sydney, Australia

Cross-Lingual Consistency Indian Languages

Ask PDF