Cross-Lingual Sentiment Misalignment: Auditing Multilingual Language Models for Inversion Risk, Dialectal Representation, and Affective Stability

Cross-Lingual Sentiment Misalignment: Auditing Multilingual Language
Models for Inversion Risk, Dialectal Representation, and Affective Stability
Nusrat Jahan Lia
Institute of Information Technology
University of Dhaka, Dhaka, Bangladesh
bsse1306@iit.du.ac.bd
Shubhashis Roy Dipta
University of Maryland, Baltimore County
Baltimore, Maryland, USA
sroydip1@umbc.edu
Abstract
Recent advances in multilingual representa-
tion learning aim to bridge the performance
gap between high- and low-resource languages,
yet their ability to preserve affective meaning
across languages remains underexplored, par-
ticularly for underrepresented languages like
Bengali. This research addresses cross-lingual
sentiment misalignment between Bengali and
English by introducing a controlled benchmark-
ing framework evaluating four multilingual
transformer models on parallel Bengali-English
sentence pairs, stratified by dialect, to assess
their representational stability. We demonstrate
that a compressed model architecture exhibits
a 28.7% “Sentiment Inversion Rate,” funda-
mentally misinterpreting positive semantics as
negative (or vice versa). Consequently, we
identify a cross-lingual sentiment skew that
we call “Asymmetric Empathy”, where models
systematically dampen or artificially amplify
the affective weight of Bengali text relative to
its exact English counterpart. Finally, we ex-
pose a key vulnerability regarding dialectal rep-
resentation: a “Modern Bias” in the regional
model, which exhibits a 57% increase in align-
ment error when processing the formal Bengali
register compared to modern colloquial text.
As foundational encoders continue to serve as
safety classifiers and reward models for LLM
pipelines, cross-lingual reliability becomes a
critical concern. We therefore advocate for the
integration of “Affective Stability” metrics into
future cross-lingual benchmarks to detect and
penalize polarity inversions, particularly in low-
resource settings.
1 Introduction
Multilingual

iers and reward models for LLM
pipelines, cross-lingual reliability becomes a
critical concern. We therefore advocate for the
integration of “Affective Stability” metrics into
future cross-lingual benchmarks to detect and
penalize polarity inversions, particularly in low-
resource settings.
1 Introduction
Multilingual models have rapidly become the
backbone of language-agnostic information ac-
cess, spanning sentiment analysis, content modera-
tion, retrieval-augmented systems, and downstream
knowledge-intensive applications. However, when
Negative	Neutral	Positive	Negative
Neutral
Positive
Alignment Corridor (Y=X)
All that day, his insulted and angry heart began to rule his instincts
সিদন 	সমস্ত 	িদন 	ধিরয়া 	তাহার 	অপ
মানাহত 	ুব্ধ 	িচত্ত 	তাহার প্র 	বৃিত্ত 	েক 	শাসন 	কিরেত 	লািগল
Figure 1: The plot maps the predicted English sentiment
vector (X-axis) against the Bengali vector (Y -axis) for
a test sentence. The green corridor represents ideal
cross-lingual sentiment alignment (Y ≈ X). While
the large-scale XLM-T model successfully preserves
polarity, the compressed (mDistilBERT) and regionally
specialized (IndicBERT) architectures are projecting a
negative English statement into a positive Bengali latent
space (Red Zone).
a multilingual model correctly identifies that an En-
glish sentence carries negative sentiment but simul-
taneously classifies its Bengali semantic equivalent
as positive (Figure 1), the representational promise
of multilingualism collapses in practice.
Multilingual sentiment encoders, including those
for Bengali, are deployed as core components in
content moderation systems, where biases can prop-
agate to real-world safety failures; for instance, an
audit of Bengali sentiment tools revealed identity-
based inconsistencies that undermine moderation
accuracy (Das et al., 2024). Similarly, Tan et al.
(2025) use multilingual classifiers for government-
scale moderation in diverse languages with reliance
on foundational encoders for LLM safety

ilures; for instance, an
audit of Bengali sentiment tools revealed identity-
based inconsistencies that undermine moderation
accuracy (Das et al., 2024). Similarly, Tan et al.
(2025) use multilingual classifiers for government-
scale moderation in diverse languages with reliance
on foundational encoders for LLM safety pipelines.
Such instances illustrate the risks of sentiment mis-
alignment in operational safety classifiers.
1
arXiv:2602.17469v2 [cs.CL] 30 Apr 2026

-- 1 of 10 --

To mitigate the curse of multilinguality, the phe-
nomenon where adding more languages to a fixed-
capacity model degrades per-language performance
(Conneau et al., 2020), we investigate the “Senti-
ment Inversion” crisis and its broader representa-
tional implications. We demonstrate that the multi-
lingual curse manifests not only as accuracy degra-
dation but also as affective inversion. We employ
a standardized cross-lingual sentiment alignment
framework to evaluate the semantic consistency of
four multilingual sentiment classifier models on
parallel Bengali–English text (see Table 4).
We show that alignment is not uniform across
languages or dialects, as summarized in the follow-
ing key findings:
Capacity Constraints & Alignment Instabil-
ity: We demonstrate that the capacity-constrained
mDistilBERT architecture exhibits a 28.7% “Senti-
ment Inversion Rate,” accompanied by a heavy-
tailed distribution of errors (Figure 3), suggest-
ing that model compression may compromise the
safety margins required for cross-lingual alignment.
We further find that regional specialization alone
does not resolve this issue; however, a distilled ar-
chitecture (Tabularis) that leverages synthetic data
can reduce such representational misalignments.
Asymmetric Empathy: We reveal that mod-
els exhibit consistent cross-lingual directional bias,
systematically dampening or artificially inflating
the emotional intensity of non-English text.
The Dialectal Sentiment Gap: We find a “Mod-
ern Bias” in which

s synthetic data
can reduce such representational misalignments.
Asymmetric Empathy: We reveal that mod-
els exhibit consistent cross-lingual directional bias,
systematically dampening or artificially inflating
the emotional intensity of non-English text.
The Dialectal Sentiment Gap: We find a “Mod-
ern Bias” in which models align well with col-
loquial text but underperform on formal dialects,
which are the backbone of Bengali literature.
2 Related Works
This section reviews prior work in multilingual rep-
resentation, cross-lingual sentiment analysis, and
dialect-aware NLP, with a focus on gaps in affective
alignment and evaluation.
2.1 Bengali in the Multilingual NLP
Landscape
Bengali remains one of the most underrepresented
major languages in NLP despite its global speaker
base. Kabir et al. (2024) provide a comprehen-
sive audit of LLM performance on Bengali NLP
tasks, identifying key failure modes including lan-
guage generation errors, verbose mismatching with
evaluation metrics, and task-specific weaknesses.
Bhowmik et al. (2025) document consistent perfor-
mance gaps for Bengali relative to English across
recent LLMs, tracing these to tokenization ineffi-
ciency, where models fragment Bengali script into
excessive subword units, degrading semantic coher-
ence. The BnMMLU benchmark further demon-
strates that even large-scale frontier models show
sublinear returns in Bengali reasoning as model
size increases, an empirical fingerprint of the mul-
tilingual curse at scale (Joy and Shatabda, 2025;
Conneau et al., 2020).
For sentiment analysis specifically, BanglaBERT
(Bhattacharjee et al., 2022) established a strong
monolingual baseline, and ensemble transformer
systems (Hoque et al., 2024) have achieved high ag-
gregate accuracy. However, in a landmark audit of
Bengali sentiment analysis tools, Das et al. (2024)
reveal that aggregate accuracy conceals systematic
identity-based biases, with tools exhibiting differ-
ential performance across gender, religious,

mble transformer
systems (Hoque et al., 2024) have achieved high ag-
gregate accuracy. However, in a landmark audit of
Bengali sentiment analysis tools, Das et al. (2024)
reveal that aggregate accuracy conceals systematic
identity-based biases, with tools exhibiting differ-
ential performance across gender, religious, and
national identity signals. This colonial impulse in
tool design highlights how reductionist representa-
tions reanimate historical hierarchies and motivates
the external auditing approach we adopt. Our work
extends this critical perspective by focusing specif-
ically on cross-lingual affective misalignment and
dialectal representational harm.
2.2 The Multilingual Curse and Capacity
Constraints
The curse of multilinguality, originally formalized
by Conneau et al. (2020), describes the empiri-
cal observation that, under fixed model capacity,
adding more languages to pretraining initially ben-
efits low-resource languages through positive trans-
fer but eventually degrades per-language perfor-
mance due to inter-language parameter competi-
tion. Recent work has substantially refined this
understanding. Blevins et al. (2024) demonstrate
that the curse can be partially lifted through Cross-
lingual Expert Language Models (X-ELM), which
decouple per-language capacity via modular train-
ing and outperform jointly trained multilingual
models across 16 languages. Foroutan et al. (2025)
further argue that the curse arises not from language
count per se but from finite model capacity ampli-
fying the impact of noisy, low-quality data in low-
resource languages. This has direct implications
for Bengali, where pretraining data is both scarce
and noisier than for high-resource languages.
2

-- 2 of 10 --

2.3 Cross-Lingual Sentiment Analysis and
Representational Failures
The adoption of transformer architectures has
substantially advanced sentiment analysis in low-
resource languages (Bhowmick and Jana, 2021).
Cross-lingual transfer from high-resource to

e
and noisier than for high-resource languages.
2

-- 2 of 10 --

2.3 Cross-Lingual Sentiment Analysis and
Representational Failures
The adoption of transformer architectures has
substantially advanced sentiment analysis in low-
resource languages (Bhowmick and Jana, 2021).
Cross-lingual transfer from high-resource to low-
resource languages has been enabled both through
shared representations in pretrained multilingual
models (Conneau et al., 2020) and through machine
translation strategies (Poncelas et al., 2020). Chen
et al. (2025) document that while GPT-4 achieves
approximately 84.4% F1 in English sentiment, this
drops to around 67% for low-resource languages,
and propose adaptive self-alignment strategies with
data augmentation to partially close this gap. Re-
cent literature shows that hybrid approaches retain-
ing lexicon features maintain stability advantages
over purely neural representations (Mahmud et al.,
2024), and ensemble methods can achieve high
aggregate accuracy (Hoque et al., 2024).
A growing body of literature documents that
high aggregate accuracy routinely masks represen-
tational failures (Das et al., 2024). Wasi et al.
(2024) show that LLMs acquire social biases
through surface linguistic cues, and that Bengali di-
alectal variation, particularly religious dialect vari-
ation, induces systematic performance divergence
in large models. Ochieng et al. (2025) extend this
critique, demonstrating that reasoning-based LLM
sentiment evaluation in low-resource, culturally
nuanced contexts reveals failures invisible to label-
prediction benchmarks. The CuLEmo benchmark
(Belay et al., 2025) further shows that multilin-
gual LLMs systematically fail to capture culturally
grounded variations in emotional expression across
languages. These cultural layers of failure are dis-
tinct from, but related to, the cross-lingual affective
misalignment we document.
2.4 Bengali Diglossia and Dialectal NLP
Dialectal variation represents a significant chal-
lenge

matically fail to capture culturally
grounded variations in emotional expression across
languages. These cultural layers of failure are dis-
tinct from, but related to, the cross-lingual affective
misalignment we document.
2.4 Bengali Diglossia and Dialectal NLP
Dialectal variation represents a significant chal-
lenge for multilingual NLP, as Wasi et al. (2024)
demonstrate through empirical evaluation of LLMs
on Bengali religious dialects. Bengali exhibits a
well-documented diglossic structure comprising
Sadhu Bhasha (formal/literary, Sanskrit-derived vo-
cabulary, archaic conjugation) and Cholito Bhasha
(colloquial/standard, simplified morphology, con-
temporary vocabulary), a stylistic split that presents
significant challenges for multilingual NLP due to
the frequent blending of these forms in everyday
communication (Ayman et al., 2025). The critical
insight motivating our dialect stratification is that
training data for multilingual models is overwhelm-
ingly drawn from contemporary digital sources
such as large-scale web crawls, social media, and
news corpora (Conneau et al., 2020; Kakwani et al.,
2020), creating a training distribution that is inher-
ently skewed toward the more common, modern
Cholito Bengali form (Ayman et al., 2025).
2.5 Benchmarking Gaps and the Need for
Affective Stability Metrics
Current multilingual benchmarks, including
XTREME, XNLI, and their derivatives, evaluate
cross-lingual performance on semantic tasks such
as natural language inference, question answering
(e.g., MLQA, TyDiQA), and named entity recog-
nition (e.g., WikiAnn), with XNLI focusing on en-
tailment classification (Hu et al., 2020). While ef-
fective at measuring semantic transfer, these bench-
marks largely overlook affective fidelity. Recent
efforts such as MMAFFBen (Liu et al., 2025) begin
to address this gap by introducing affective eval-
uation, but comprehensive measurement of senti-
ment preservation and cross-lingual affective align-
ment remains limited.

at measuring semantic transfer, these bench-
marks largely overlook affective fidelity. Recent
efforts such as MMAFFBen (Liu et al., 2025) begin
to address this gap by introducing affective eval-
uation, but comprehensive measurement of senti-
ment preservation and cross-lingual affective align-
ment remains limited. Ochieng et al. (2025) ex-
plicitly call for benchmarks that measure LLM
sentiment in low-resource, culturally nuanced con-
texts beyond label accuracy. Miah et al. (2024)
note that translation-based cross-lingual sentiment
approaches can achieve high aggregate accuracy
while failing to capture culturally grounded vari-
ations in emotional expression, often introducing
translation biases in affective intensity.
We interpret these findings as evidence of di-
rectional distortions in how sentiment is mapped
across languages, which we formalize as asym-
metric empathy. Recent work on multilingual bias
evaluation (Wasi et al., 2024; Sadhu et al., 2025)
has established that social bias in Bengali LLMs op-
erates across gender and religious lines, but has not
examined the cross-lingual affective alignment di-
mension we investigate. Our proposal for affective
stability metrics, which explicitly penalize polar-
ity inversions and dialectal divergence, responds
to this benchmarking gap, extending recent work
on stability-focused evaluation metrics (Atil et al.,
2024) to the multilingual affective domain.
3

-- 3 of 10 --

3 Methodology
We employ a controlled experimental framework
to quantify cross-lingual sentiment alignment in
multilingual transformer architectures. We adopt
a within-model comparative design in which each
transformer processes parallel Bengali-English text
pairs independently, enabling direct measurement
of semantic divergence without inter-model archi-
tectural confounds.
3.1 Dataset Specification
We utilize a parallel corpus comprising n = 7,350
Bengali-English sentence pairs, sourced from the
publicly available “BanglaBlend” dataset

sses parallel Bengali-English text
pairs independently, enabling direct measurement
of semantic divergence without inter-model archi-
tectural confounds.
3.1 Dataset Specification
We utilize a parallel corpus comprising n = 7,350
Bengali-English sentence pairs, sourced from the
publicly available “BanglaBlend” dataset (Ayman
et al., 2025). Formally, the dataset D is defined as
a set of tuples:
D = {(Bi, Ei, Di) | i ∈ [1, n]} (1)
where:
• Bi ∈ Σ∗
Bengali represents the i-th Bengali sen-
tence (original text),
• Ei ∈ Σ∗
English represents the corresponding
English translation,
• Di ∈ {Sadhu, Cholito} denotes the Bengali
dialect classification.
Bengali exhibits diglossia with two primary writ-
ten forms:
1. Sadhu Bhasha: Formal/literary register, char-
acterized by Sanskrit-derived vocabulary and
archaic verb conjugations.
2. Cholito Bhasha: Colloquial/standard regis-
ter with relatively simplified morphology and
contemporary vocabulary.
3.2 Models Evaluated
We benchmark four multilingual transformer ar-
chitectures representing distinct design paradigms:
XLM-T, a large-scale model fine-tuned on high-
volume multilingual social media data; IndicBERT,
a regionally specialized encoder for Indian lan-
guages; Tabularis, a distilled multilingual model en-
hanced with synthetic data for broad cross-lingual
coverage; and mDistilBERT, a compressed multi-
lingual model designed for efficient zero-shot sen-
timent transfer. Full repository mappings are pro-
vided in Table 4. This selection enables systematic
analysis of how scale, regional specialization, syn-
thetic augmentation, and compression interact with
cross-lingual affective alignment.
These specific models were selected based on
three core inclusion criteria: they are publicly ac-
cessible, they are capable of zero-shot Bengali
sentiment inference without requiring further task-
specific fine-tuning, and together they cover a prin-
cipled spectrum of architectural capacity (large-
scale vs. compressed) and design

models were selected based on
three core inclusion criteria: they are publicly ac-
cessible, they are capable of zero-shot Bengali
sentiment inference without requiring further task-
specific fine-tuning, and together they cover a prin-
cipled spectrum of architectural capacity (large-
scale vs. compressed) and design intent (global vs.
regional). Furthermore, because our experimental
design isolates cross-lingual representation and af-
fective transfer as the primary variables of interest,
we focus exclusively on multilingual architectures,
purposefully excluding monolingual models.
3.3 Experimental Design
For each model M and sentence pair (Bi, Ei), we
perform independent inference on both languages
using the same model weights θM :
ˆyB,i = Mθ(Bi) [Bengali stream] (2)
ˆyE,i = Mθ(Ei) [English stream] (3)
Any divergence between ˆyB,i and ˆyE,i is at-
tributable to cross-lingual representation, calibra-
tion, or decision boundary alignment within the
same parameter space.
3.4 Score Normalization and Metric
Formulation
3.4.1 Universal Score Normalizer
To enable direct comparison, we define a universal
normalization function φ : Predictions → [−1, 1]:
φ(ˆy) =

	
	
ˆy.score if ˆy.label ∈ {positive}
−ˆy.score if ˆy.label ∈ {negative}
0 if ˆy.label ∈ {neutral}
(4)
*Note: Standard “Positive” and “Negative” labels
map to ±s for 2-class and 3-class models, while
intermediate positive/negative classes are scaled by
0.5 in the 5-class Tabularis model to accommodate
the “Very Positive” and “Very Negative” extremes.
This maintains a uniform linear spacing across
sentiment intensity levels, ensuring that the five
classes are equidistant on the sentiment continuum.
4

-- 4 of 10 --

After normalization, we obtain continuous senti-
ment scores:
SB,i = φ(ˆyB,i) ∈ [−1, 1] (5)
SE,i = φ(ˆyE,i) ∈ [−1, 1] (6)
where SB,i and SE,i denote the Bengali and En-
glish sentiment scores, respectively.
3.4.2 Sentence-Level Alignment Metrics
For each sentence pair i, we compute four

nt continuum.
4

-- 4 of 10 --

After normalization, we obtain continuous senti-
ment scores:
SB,i = φ(ˆyB,i) ∈ [−1, 1] (5)
SE,i = φ(ˆyE,i) ∈ [−1, 1] (6)
where SB,i and SE,i denote the Bengali and En-
glish sentiment scores, respectively.
3.4.2 Sentence-Level Alignment Metrics
For each sentence pair i, we compute four align-
ment metrics:
M1. Alignment Divergence
Di = |SB,i − SE,i| ∈ [0, 2] (7)
Interpretation: Di = 0.0 indicates perfect align-
ment (identical sentiment), Di = 0.5 indicates
moderate divergence, Di = 2.0 indicates maximal
divergence (opposite extremes).
M2. Directional Bias
Bi = SE,i − SB,i ∈ [−2, 2] (8)
Interpretation: Bi > 0 indicates that the English
text is predicted as more positive (or less negative)
than its Bengali counterpart; Bi < 0 indicates the
reverse; Bi ≈ 0 indicates minimal cross-lingual
divergence for that pair.
M3. Polarity Inversion (Safety Metric)
Ii = 1
"(SB,i > τ ∧ SE,i < −τ )
∨ (SB,i < −τ ∧ SE,i > τ )
#
(9)
where 1[·] is the indicator function and τ = 0.1
is a noise threshold to avoid false positives from
near-zero scores.
Interpretation: Ii = 1 indicates sentiment in-
version (e.g., Bengali=Positive, English=Negative);
Ii = 0 indicates polarity preserved. Inversion is
the most severe failure mode, as it indicates mis-
alignment of sentiment direction.
3.5 Population-Level Aggregation and
Statistical Computation
To characterize model-level performance and eval-
uate representational equity across Bengali diglos-
sia, we aggregate the sentence-level metrics into
population-level statistics. We compute these
across the entire dataset D as well as its stratified
dialectal subsets (DSadhu and DCholito).
Metric 	Tabularis XLM-T IndicBERT mDistilBERT
Mean Div. 	0.200 0.276 	0.375 	0.417
Std Dev. 	0.214 0.298 	0.607 	0.429
Sadhu Div. 	0.239 0.286 	0.459 	0.456
Cholito Div. 	0.161 0.266 	0.292 	0.379
Dialect Gap 	0.078 0.020 	0.167 	0.077
Sadhu Err. Inc. (%) 48.4 	7.6 	57.1 	20.5
Robustness (%) 	43.1 	42.1 	58.3 	34.2
Inversions 	635

aris XLM-T IndicBERT mDistilBERT
Mean Div. 	0.200 0.276 	0.375 	0.417
Std Dev. 	0.214 0.298 	0.607 	0.429
Sadhu Div. 	0.239 0.286 	0.459 	0.456
Cholito Div. 	0.161 0.266 	0.292 	0.379
Dialect Gap 	0.078 0.020 	0.167 	0.077
Sadhu Err. Inc. (%) 48.4 	7.6 	57.1 	20.5
Robustness (%) 	43.1 	42.1 	58.3 	34.2
Inversions 	635 	267 	1471 	2107
Inv. Rate (%) 	8.6 	3.6 	20.0 	28.7
Dir. Bias (En-Bn) 0.002 0.057 	0.106 	-0.066
Table 1: Metric scores per model
Table 2 details the mathematical formulations
and interpretations for all population-level evalua-
tion criteria, including overall alignment statistics,
safety indicators, and specialized metrics designed
to quantify the dialectal gap.
4 Results
Beyond average alignment error, our evaluation
identifies three critical failure modes in multilin-
gual model behavior. Table 1 presents the corre-
sponding quantitative results.
4.1 Finding 1: Sentiment Inversion and
Alignment Instability
We define a sentiment inversion as a case where
a Bengali-English translation pair (similar mean-
ing) receives opposite polarity classifications (pos-
itive vs. negative). Such inversions represent ma-
jor alignment failures, as propositional meaning is
preserved but affective interpretation is reversed.
Across model architectures, inversion rates vary
dramatically, revealing a potential relationship be-
tween compression, divergence magnitude, and op-
timization (see Figure 2).
Figure 2: Sentiment Inversion Rate Across Models
• Compression and Elevated Inversion Risk.
The distilled multilingual architecture (mDis-
tilBERT) exhibits the highest mean alignment
divergence and lowest robustness (see Ta-
ble 1). Nearly one in three sentence pairs
5

-- 5 of 10 --

Metric Formulation Interpretation
Mean Alignment Error
(Divergence)
μD (M ) = 1
n
Pn
i=1 Di Lower μD indicates better average cross-
lingual consistency.
Std. Deviation
(Divergence)
σD (M ) =
q 1
n−1
Pn
i=1(Di − μD )2 Quantifies the variability in alignment qual-
ity across the

n three sentence pairs
5

-- 5 of 10 --

Metric Formulation Interpretation
Mean Alignment Error
(Divergence)
μD (M ) = 1
n
Pn
i=1 Di Lower μD indicates better average cross-
lingual consistency.
Std. Deviation
(Divergence)
σD (M ) =
q 1
n−1
Pn
i=1(Di − μD )2 Quantifies the variability in alignment qual-
ity across the dataset.
Robustness Index R(M ) = 1
n
Pn
i=1 1[Di < 0.1] × 100% % of pairs with negligible divergence; mea-
sures the “safe operating zone.”
Inversion Rate Irate(M ) = 1
n
Pn
i=1 Ii × 100% % of sentence pairs exhibiting sentiment
polarity flips.
Mean Directional Bias μB (M ) = 1
n
Pn
i=1 Bi > 0: English favored; < 0: Bengali fa-
vored; ≈ 0: No systematic language skew.
Formal Penalty
(Dialect Gap)
∆dialect(M ) = μD (M, Sadhu) − μD (M, Cholito) > 0 implies a “Modern Bias” where the
model struggles with formal registers.
Relative Dialect Error ∆%
dialect(M ) = ∆dialect(M )
μD (M,Cholito) × 100% Normalizes the formal penalty by the base-
line colloquial error rate for fair compari-
son.
Table 2: Summary of population-level alignment and dialectal metrics.
processed by the compressed model receives
directly contradictory affective classifications
across Bengali and English. This indicates
that while compression improves efficiency, it
may disproportionately reduce the representa-
tional capacity required for reliable affective
calibration. As a result, sentiment polarity re-
versal emerges as a critical failure mode that
distorts core cross-lingual meaning.
• Heavy-Tailed Failure Distribution. Align-
ment error density analysis (Figure 3) shows
that divergence is not normally distributed. In-
stead, the compressed architecture exhibits a
long right tail, corresponding to extreme po-
larity flips. While IndicBERT achieves the
highest robustness metric, it simultaneously
exhibits a massive standard deviation in diver-
gence and a high inversion rate with alignment
errors that are not normally distributed. This
non-normal distribution has practical

its a
long right tail, corresponding to extreme po-
larity flips. While IndicBERT achieves the
highest robustness metric, it simultaneously
exhibits a massive standard deviation in diver-
gence and a high inversion rate with alignment
errors that are not normally distributed. This
non-normal distribution has practical implica-
tions: mean divergence understates actual risk,
and models cannot be reliably characterized
by their average behavior for deployment in
critical downstream applications.
• Scale-Driven Inversion Resilience. The
large-scale multilingual model (XLM-T)
records the lowest polarity inversion rate (Ta-
ble 1) across all evaluation pairs. This sug-
gests that massive parameter scale and diverse
pre-training may preserve coherent affect map-
pings more effectively than regional special-
ization, buffering against semantic instability.
Crucially, however, the robust distilled Tabu-
laris model (8.6% inversion rate) implies that
compression with data-centric optimization
can substantially close the gap between com-
pressed and full-scale architectures. Hence,
scale is not the only path to alignment stability.
As a DistilBERT-based model fine-tuned with
diverse synthetic multilingual data, Tabularis
shows that targeted training strategies, rather
than scale alone, can drive alignment stability.
Figure 3: Distribution of Alignment Error Density
4.2 Finding 2: Representational Harm and
the Dialectal Gap
To evaluate robustness under Bengali diglossia, we
compare alignment divergence across colloquial
(Cholito) and formal (Sadhu) variants. A multilin-
gual system should maintain stable cross-lingual
6

-- 6 of 10 --

calibration regardless of lexical register. However,
we observed a dialectal sensitivity pattern.
• Modern-Register Overfitting. Both In-
dicBERT and Tabularis exhibit sharp in-
creases in divergence when processing for-
mal Sadhu text (see Table 1). This indicates
that alignment quality is concentrated in mod-
ern and high-frequency lexical

f lexical register. However,
we observed a dialectal sensitivity pattern.
• Modern-Register Overfitting. Both In-
dicBERT and Tabularis exhibit sharp in-
creases in divergence when processing for-
mal Sadhu text (see Table 1). This indicates
that alignment quality is concentrated in mod-
ern and high-frequency lexical distributions,
while archaic or formal constructions fall out-
side the model’s calibrated semantic mani-
fold. We term this phenomenon Modern Bias:
strong alignment in contemporary usage, but
underperformance in formal registers. The
Modern Bias finding has direct consequences
for linguistic equity: users who communi-
cate in formal registers including academic,
literary, and administrative Bengali, receive
poorer cross-lingual sentiment alignment than
users of colloquial varieties. As a result, cur-
rent multilingual systems risk institutionaliz-
ing a structural inequity in which access to
reliable inference is contingent on conforming
to simplified or non-native linguistic norms.
Conversely, XLM-T demonstrates strong di-
alectal resilience, likely due to its large-scale
multilingual training over diverse text distri-
butions, which provides broader lexical and
syntactic coverage and supports stable affec-
tive mappings across regional variation.
4.3 Finding 3: Asymmetric Empathy and
Directional Bias
In a multilingual architecture, the system must pre-
serve not just the polarity but the intensity of user
intent. We evaluate this using Directional Bias (En-
glish score − Bengali score). An ideally aligned
system should yield a distribution centered at zero
with low variance. Instead, we observed two dis-
tinct alignment regimes (as seen in Table 1, Fig-
ure 4 and Figure 5)
• Compression-Induced Bengali Positivity
Skew. The distilled architecture (mDistil-
BERT) exhibits a negative directional bias, in-
dicating that it systematically scores Bengali
text as more positive (or less negative) than
its exact English translation. For a human
user in a

seen in Table 1, Fig-
ure 4 and Figure 5)
• Compression-Induced Bengali Positivity
Skew. The distilled architecture (mDistil-
BERT) exhibits a negative directional bias, in-
dicating that it systematically scores Bengali
text as more positive (or less negative) than
its exact English translation. For a human
user in a safety-critical context: the model
may artificially dampen the severity of a neg-
ative Bengali sentiment, underweighting the
severity of negative Bengali content relative to
equivalent English content. As demonstrated
in the case study (Figure 5), mDistilBERT ex-
hibits a safety failure by correctly scoring an
English statement as deeply negative (-0.981)
while assigning a positive score (+0.533) for
its exact Bengali equivalent.
• Regional English Optimism Bias. Con-
versely, IndicBERT demonstrates a positive
mean directional bias, assigning higher posi-
tivity (or lower negativity) to English inputs
relative to their Bengali counterparts. This
imposes an equity penalty on Bengali users,
as their neutral or moderately negative state-
ments are penalized with harsher negative clas-
sifications compared to English speakers.
These findings indicate that cross-lingual affec-
tive misalignment is consistently and directionally
structured rather than stochastic, which can lead to
downstream distortions in applications relying on
Bengali sentiment signals.
4.4 Proposed Metric: Affective Stability Index
Our empirical evaluation reveals that aggregate
alignment error (μD) alone is insufficient to capture
the critical safety failures inherent in cross-lingual
representation. Specifically, models with relatively
moderate average divergence may still exhibit high
rates of sentiment inversion (Irate) (Table 1). To
address this benchmarking gap and formally quan-
tify cross-lingual reliability, we introduce the Af-
fective Stability Index (AS). We define Affective
Stability as a composite metric that rewards tight se-
mantic alignment while strictly penalizing

still exhibit high
rates of sentiment inversion (Irate) (Table 1). To
address this benchmarking gap and formally quan-
tify cross-lingual reliability, we introduce the Af-
fective Stability Index (AS). We define Affective
Stability as a composite metric that rewards tight se-
mantic alignment while strictly penalizing polarity
inversions. Utilizing the population-level metrics
defined in Section 3.5, it is computed as:
AS(M ) =

1 − μD(M )
2

× (1 − Irate(M ))
(10)
The first term normalizes the Mean Alignment
Divergence (μD) into a similarity score bounded
by [0, 1], as the maximum theoretical divergence
in our normalized space is 2.0. The second term
acts as a strict penalty mask based on the Inversion
Rate (Irate), expressed as a probability. An AS
score of 1.0 indicates perfect cross-lingual affective
fidelity, whereas lower scores reflect compounding
representational misalignments.
7

-- 7 of 10 --

Table 3 presents the Affective Stability scores
for all evaluated architectures. The results validate
our observations regarding scale and compression.
While the large-scale XLM-T maintains high affec-
tive stability, mDistilBERT suffers a degradation
in overall reliability. Notably, the distilled Tabu-
laris model achieves an AS score highly compet-
itive with XLM-T. This empirically demonstrates
that targeted training strategies and synthetic data
utilization can preserve affective calibration even
under capacity constraints.
Model 	Mean Div. (μD) Inv. Rate (Irate) Affective Stability (AS)
XLM-T 	0.276 	0.036 	0.831
Tabularis 	0.200 	0.086 	0.823
IndicBERT 	0.375 	0.200 	0.650
mDistilBERT 	0.417 	0.287 	0.564
Table 3: Affective Stability (AS) of evaluated models.
5 Discussion: Alignment Under
Linguistic Pluralism
Our findings instantiate the multilingual curse at an
affective, rather than purely accuracy-based, level.
The inversion gap we observe is consistent with
Blevins et al. (2024)’s theoretical framing: fixed
model capacity induces inter-language

AS) of evaluated models.
5 Discussion: Alignment Under
Linguistic Pluralism
Our findings instantiate the multilingual curse at an
affective, rather than purely accuracy-based, level.
The inversion gap we observe is consistent with
Blevins et al. (2024)’s theoretical framing: fixed
model capacity induces inter-language parameter
competition that degrades per-language represen-
tation. We hypothesize that compression amplifies
this competition in ways that may disrupt the fine-
grained representational signals required for stable
sentiment polarity mapping.
Our results suggest two distinct intervention
points for multilingual model development. First,
at the pretraining data level: the Modern Bias find-
ing points to a gap in training corpus composi-
tion for Bengali, even when utilizing synthetic
data. Formal Sadhu text is underrepresented in
web-crawled multilingual corpora (dominated by
social media and news in contemporary registers),
and correcting this imbalance would directly ad-
dress dialectal representational harm. Second, at
the post-distillation fine-tuning level: our finding
that Tabularis substantially outperforms mDistil-
BERT despite both being distilled architectures
demonstrates that targeted fine-tuning with syn-
thetic multilingual corpora can preserve affective
calibration through compression. Addressing both
stages is particularly vital in low-resource settings
like Bengali, where computationally efficient, com-
pressed multilingual models must still deliver eq-
uitable and reliable affective understanding across
all linguistic communities.
The multilingual NLP evaluation landscape cur-
rently lacks standardized metrics for cross-lingual
affective consistency. Existing benchmarks (Hu
et al., 2020; Ruder et al., 2021; Han et al., 2025;
Goldman et al., 2025) primarily measure semantic
and syntactic transfer but do not penalize polarity
inversions or account for affective fidelity. We pro-
pose that multilingual benchmarks incorporate the
metric

-lingual
affective consistency. Existing benchmarks (Hu
et al., 2020; Ruder et al., 2021; Han et al., 2025;
Goldman et al., 2025) primarily measure semantic
and syntactic transfer but do not penalize polarity
inversions or account for affective fidelity. We pro-
pose that multilingual benchmarks incorporate the
metric suite introduced in this work (Inversion Rate,
Affective Stability) as standard evaluation dimen-
sions. This is especially critical for downstream ap-
plications that depend on affective signals: content
moderation, mental health monitoring, customer
feedback analysis, and social listening systems op-
erating across languages all face systematic failure
risks due to sentiment inversion.
We further argue that the choice is not binary
between scale and alignment; targeted distillation
strategies, data augmentation, and curated metrics
that explicitly prioritize affective calibration can
simultaneously address efficiency constraints and
ensure representational equity, particularly in low-
resource deployment settings.
6 Conclusion
We present a cross-lingual sentiment alignment au-
dit comparing four transformer architectures on
Bengali-English parallel data stratified by dialect.
Our findings reveal that current multilingual mod-
els exhibit structured affective representational fail-
ures: sentiment inversion under compression, di-
alectal bias against formal registers, and direc-
tional asymmetry in emotional intensity calibra-
tion. These failures are distinct from accuracy
degradation measured by standard benchmarks and
constitute specific threats to equitable language
technology access for Bengali users. We argue
that building better multilingual representations
requires evaluative interventions that make affec-
tive alignment failures visible. Future multilingual
benchmarks should incorporate Inversion Rate and
Affective Stability as standard dimensions. Future
training and alignment research should address the
formal register gap in Bengali

tter multilingual representations
requires evaluative interventions that make affec-
tive alignment failures visible. Future multilingual
benchmarks should incorporate Inversion Rate and
Affective Stability as standard dimensions. Future
training and alignment research should address the
formal register gap in Bengali pretraining data and
explore dialect-stratified post-training alignment as
a path to equitable compression. We believe these
directions are generalizable beyond Bengali to the
broader landscape of low-resource and dialectally
diverse languages underrepresented in current mul-
tilingual NLP infrastructure.
8

-- 8 of 10 --

References
Berk Atil, Alexa Chittams, Liseng Fu, Ferhan Ture,
Lixinyu Xu, and Breck Baldwin. 2024. Llm stabil-
ity: A detailed analysis with some surprises. arXiv
preprint arXiv:2408.04667, 1.
Umme Ayman, Chayti Saha, Azmain Mahtab Rahat, and
Sharun Akter Khushbu. 2025. Banglablend: A large-
scale nobel dataset of bangla sentences categorized
by saint and common form of bangla language. Data
in Brief, 58:111240.
Francesco Barbieri, Luis Espinosa Anke, and Jose
Camacho-Collados. 2022. Xlm-t: Multilingual lan-
guage models in twitter for sentiment analysis and
beyond. In Proceedings of the thirteenth language
resources and evaluation conference, pages 258–266.
Tadesse Destaw Belay, Ahmed Haj Ahmed, Alvin
C Grissom Ii, Iqra Ameer, Grigori Sidorov, Olga
Kolesnikova, and Seid Muhie Yimam. 2025. Culemo:
Cultural lenses on emotion-benchmarking llms for
cross-cultural emotion understanding. In Proceed-
ings of the 63rd Annual Meeting of the Association
for Computational Linguistics (Volume 1: Long Pa-
pers), pages 18894–18909.
Abhik Bhattacharjee, Tahmid Hasan, Wasi Ahmad,
Kazi Samin Mubasshir, Md Saiful Islam, Anindya
Iqbal, M Sohel Rahman, and Rifat Shahriyar. 2022.
Banglabert: Language model pretraining and bench-
marks for low-resource language understanding eval-
uation in bangla. In Findings of the Association
for Computational

18909.
Abhik Bhattacharjee, Tahmid Hasan, Wasi Ahmad,
Kazi Samin Mubasshir, Md Saiful Islam, Anindya
Iqbal, M Sohel Rahman, and Rifat Shahriyar. 2022.
Banglabert: Language model pretraining and bench-
marks for low-resource language understanding eval-
uation in bangla. In Findings of the Association
for Computational Linguistics: NAACL 2022, pages
1318–1327.
Anirban Bhowmick and Abhik Jana. 2021. Sentiment
analysis for bengali using transformer based models.
In Proceedings of the 18th International Conference
on Natural Language Processing (ICON), pages 481–
486.
Shimanto Bhowmik, Tawsif Tashwar Dipto, Md Saz-
zad Islam, Sheryl Hsu, and Tahsin Reasat. 2025.
Evaluating llms’ multilingual capabilities for ben-
gali: Benchmark creation and performance analysis.
arXiv preprint arXiv:2507.23248.
Terra Blevins, Tomasz Limisiewicz, Suchin Gururan-
gan, Margaret Li, Hila Gonen, Noah A Smith, and
Luke Zettlemoyer. 2024. Breaking the curse of multi-
linguality with cross-lingual expert language models.
In Proceedings of the 2024 conference on empiri-
cal methods in natural language processing, pages
10822–10837.
Vadim Borisov, Samuel Gyamfi, and Richard H.
Schreiber. 2025. Multilingual sentiment analysis.
Revision 69afb83.
Li Chen, Shifeng Shang, and Yawen Wang. 2025. Bridg-
ing resource gaps in cross-lingual sentiment anal-
ysis: adaptive self-alignment with data augmenta-
tion and transfer learning. PeerJ Computer Science,
11:e2851.
Alexis Conneau, Kartikay Khandelwal, Naman Goyal,
Vishrav Chaudhary, Guillaume Wenzek, Francisco
Guzmán, Edouard Grave, Myle Ott, Luke Zettle-
moyer, and Veselin Stoyanov. 2020. Unsupervised
cross-lingual representation learning at scale. In Pro-
ceedings of the 58th annual meeting of the associa-
tion for computational linguistics, pages 8440–8451.
Dipto Das, Shion Guha, Jed R Brubaker, and Bryan
Semaan. 2024. The“colonial impulse" of natural
language processing: An audit of bengali sentiment
analysis tools and their identity-based biases.

g at scale. In Pro-
ceedings of the 58th annual meeting of the associa-
tion for computational linguistics, pages 8440–8451.
Dipto Das, Shion Guha, Jed R Brubaker, and Bryan
Semaan. 2024. The“colonial impulse" of natural
language processing: An audit of bengali sentiment
analysis tools and their identity-based biases. In
Proceedings of the 2024 CHI Conference on Human
Factors in Computing Systems, pages 1–18.
Negar Foroutan, Paul Teiletche, Ayush Kumar Tarun,
and Antoine Bosselut. 2025. Revisiting multilingual
data mixtures in language model pretraining. arXiv
preprint arXiv:2510.25947.
Omer Goldman, Uri Shaham, Dan Malkin, Sivan Eiger,
Avinatan Hassidim, Yossi Matias, Joshua Maynez,
Adi Mayrav Gilady, Jason Riesa, Shruti Rijhwani,
and 1 others. 2025. Eclektic: a novel challenge set
for evaluation of cross-lingual knowledge transfer.
arXiv preprint arXiv:2502.21228.
Wenhan Han, Yifan Zhang, Zhixun Chen, Binbin Liu,
Haobin Lin, Bingni Zhang, Taifeng Wang, Mykola
Pechenizkiy, Meng Fang, and Yin Zheng. 2025.
Mubench: Assessment of multilingual capabilities of
large language models across 61 languages. arXiv
preprint arXiv:2506.19468.
Md Nesarul Hoque, Umme Salma, Md Jamal Uddin,
Md Martuza Ahamad, and Sakifa Aktar. 2024. Ex-
ploring transformer models in the sentiment analysis
task for the under-resource bengali language. Natu-
ral Language Processing Journal, 8:100091.
Junjie Hu, Sebastian Ruder, Aditya Siddhant, Graham
Neubig, Orhan Firat, and Melvin Johnson. 2020.
Xtreme: A massively multilingual multi-task bench-
mark for evaluating cross-lingual generalisation. In
International conference on machine learning, pages
4411–4421. PMLR.
Saman Sarker Joy and Swakkhar Shatabda. 2025.
Bnmmlu: Measuring massive multitask lan-
guage understanding in bengali. arXiv preprint
arXiv:2505.18951.
Mohsinul Kabir, Mohammed Saidul Islam, Md Tah-
mid Rahman Laskar, Mir Tafseer Nayeem, M Saiful
Bari, and Enamul Hoque. 2024. Benllm-eval: A com-
prehensive evaluation into the

rker Joy and Swakkhar Shatabda. 2025.
Bnmmlu: Measuring massive multitask lan-
guage understanding in bengali. arXiv preprint
arXiv:2505.18951.
Mohsinul Kabir, Mohammed Saidul Islam, Md Tah-
mid Rahman Laskar, Mir Tafseer Nayeem, M Saiful
Bari, and Enamul Hoque. 2024. Benllm-eval: A com-
prehensive evaluation into the potentials and pitfalls
of large language models on bengali nlp. In Pro-
ceedings of the 2024 Joint International Conference
on Computational Linguistics, Language Resources
and Evaluation (LREC-COLING 2024), pages 2238–
2252.
Divyanshu Kakwani, Anoop Kunchukuttan, Satish
Golla, Gokul NC, Avik Bhattacharyya, Mitesh M
Khapra, and Pratyush Kumar. 2020. Indicnlpsuite:
Monolingual corpora, evaluation benchmarks and
9

-- 9 of 10 --

pre-trained multilingual language models for indian
languages. In Findings of the association for compu-
tational linguistics: EMNLP 2020, pages 4948–4961.
Zhiwei Liu, Lingfei Qian, Qianqian Xie, Jimin Huang,
Kailai Yang, and Sophia Ananiadou. 2025. Mmaff-
ben: a multilingual and multimodal affective analy-
sis benchmark for evaluating llms and vlms. arXiv
preprint arXiv:2505.24423.
Hemal Mahmud, Hasan Mahmud, and Mohammad
Rifat Ahmmad Rashid. 2024. Enhancing sen-
timent analysis in bengali texts: A hybrid ap-
proach using lexicon-based algorithm and pre-
trained language model bangla-bert. arXiv preprint
arXiv:2411.19584.
Md Saef Ullah Miah, Md Mohsin Kabir, Talha Bin
Sarwar, Mejdl Safran, Sultan Alfarhood, and Md F
Mridha. 2024. A multimodal approach to cross-
lingual sentiment analysis with ensemble of trans-
former and llm. Scientific Reports, 14(1):9603.
Millicent Ochieng, Anja Thieme, Ignatius Ezeani,
Risa Ueno, Samuel Maina, Keshet Ronen, Javier
Gonzalez, and Jacki O’Neill. 2025. Reasoning
beyond labels: Measuring llm sentiment in low-
resource, culturally nuanced contexts. arXiv preprint
arXiv:2508.04199.
Alberto Poncelas, Pintu Lohar, James Hadley, and Andy
Way. 2020. The impact of indirect machine transla-
tion on

a Ueno, Samuel Maina, Keshet Ronen, Javier
Gonzalez, and Jacki O’Neill. 2025. Reasoning
beyond labels: Measuring llm sentiment in low-
resource, culturally nuanced contexts. arXiv preprint
arXiv:2508.04199.
Alberto Poncelas, Pintu Lohar, James Hadley, and Andy
Way. 2020. The impact of indirect machine transla-
tion on sentiment classification. In Proceedings of
the 14th Conference of the Association for Machine
Translation in the Americas (Volume 1: Research
Track), pages 78–88.
Sebastian Ruder, Noah Constant, Jan Botha, Aditya Sid-
dhant, Orhan Firat, Jinlan Fu, Pengfei Liu, Junjie Hu,
Dan Garrette, Graham Neubig, and 1 others. 2021.
Xtreme-r: Towards more challenging and nuanced
multilingual evaluation. In Proceedings of the 2021
Conference on Empirical Methods in Natural Lan-
guage Processing, pages 10215–10245.
Jayanta Sadhu, Maneesha Rani Saha, and Rifat Shahri-
yar. 2025. Social bias in large language models for
bangla: An empirical study on gender and religious
bias. In Proceedings of the First Workshop on Lan-
guage Models for Low-Resource Languages, pages
204–218.
Leanne Tan, Gabriel Chua, Ziyu Ge, and Roy Ka-Wei
Lee. 2025. Lionguard 2: Building lightweight, data-
efficient & localised multilingual content moderators.
In Proceedings of the 2025 Conference on Empirical
Methods in Natural Language Processing: System
Demonstrations, pages 264–285.
Azmine Toushik Wasi, Raima Islam, Mst Rafia Islam,
Taki Hasan Rafi, and Dong-Kyu Chae. 2024. Explor-
ing bengali religious dialect biases in large language
models with evaluation perspectives. arXiv preprint
arXiv:2407.18376.
A Supplementary Figures and Tables
This appendix contains the visualizations and table
referenced in the main findings of the paper.
Model Name Repository (HuggingFace)
XLM-T cardiffnlp/XLM-Toberta-sentiment
(Barbieri et al., 2022)
IndicBERT ai4bharat/IndicBERTv2-sentiment
Tabularis tabularisai/multilingual-sentiment
mDistilBERT lxyuan/distilbert-multilingual
Table 4: Model Repository

izations and table
referenced in the main findings of the paper.
Model Name Repository (HuggingFace)
XLM-T cardiffnlp/XLM-Toberta-sentiment
(Barbieri et al., 2022)
IndicBERT ai4bharat/IndicBERTv2-sentiment
Tabularis tabularisai/multilingual-sentiment
mDistilBERT lxyuan/distilbert-multilingual
Table 4: Model Repository Mapping
Note on Tabularis: This model
is a fine-tuned version of the model
distilbert/distilbert-base-multilingual-cased
for multilingual sentiment analysis. It utilizes synthetic data
from multiple sources to achieve robust performance across
different languages and cultural contexts (Borisov et al.,
2025).
Figure 4: Directional Bias in Sentiment Scores (English
− Bengali)
Figure 5: Illustrative case study validating Asymmetric
Empathy in cross-lingual sentiment alignment, demon-
strating severe instance-level directional bias.
10

-- 10 of 10 --