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

2026-06-26 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Natural Language Processing · Depth: Expert, quick

Summary

This research addresses cross-lingual sentiment misalignment between Bengali and English by introducing a controlled benchmarking framework. It evaluates four multilingual transformer models on parallel Bengali-English sentence pairs, stratified by dialect, to assess their representational stability. The study found that a compressed model architecture exhibited a 28.7% "Sentiment Inversion Rate," fundamentally misinterpreting positive semantics as negative or vice versa. It also identified "Asymmetric Empathy," where models systematically dampen or amplify the affective weight of Bengali text relative to its English counterpart. Furthermore, a "Modern Bias" was exposed in regional models, showing a 57% increase in alignment error for formal Bengali compared to modern colloquial text. Given the use of foundational encoders as safety classifiers and reward models, cross-lingual reliability is a critical concern, especially in low-resource settings. The authors advocate for integrating "Affective Stability" metrics into future cross-lingual benchmarks.

Key takeaway

For NLP Engineers or MLOps Engineers deploying multilingual LLMs, you must audit models for cross-lingual sentiment misalignment, particularly with low-resource languages like Bengali. Your current models might exhibit a 28.7% "Sentiment Inversion Rate" or "Asymmetric Empathy," misinterpreting affective meaning. Integrate "Affective Stability" metrics into your evaluation pipelines to detect and penalize polarity inversions, especially when handling diverse dialects, to ensure reliable and safe model behavior.

Key insights

Multilingual models exhibit significant cross-lingual sentiment misalignment, inversion, and dialectal bias, especially for low-resource languages like Bengali.

Principles

• Affective meaning preservation is underexplored in multilingual models.
• Compressed model architectures can yield high sentiment inversion rates.
• Regional models may show a "Modern Bias" in dialectal representation.

Method

A controlled benchmarking framework was introduced, evaluating four multilingual transformer models on parallel Bengali-English sentence pairs, stratified by dialect, to assess representational stability.

In practice

• Integrate "Affective Stability" metrics into cross-lingual benchmarks.
• Detect and penalize polarity inversions in low-resource language settings.

Topics

• Cross-lingual Sentiment Analysis
• Multilingual Language Models
• Sentiment Inversion Rate
• Affective Stability Metrics
• Low-Resource Languages
• Bengali Language Processing
• Dialectal Representation

Best for: Research Scientist, AI Scientist, NLP Engineer, MLOps Engineer

Related on AIssential

• Zero-Shot to Full-Resource: Cross-lingual Transfer Strategies for Aspect-Based Sentiment Analysis — Cross-lingual transfer strategies for ABSA vary in effectiveness based on model architecture and resource availability.
• Fully Automated Identification of Lexical Alignment and Preference-Stage Shifts in Large Language Models — Automated metrics can identify and quantify lexical overuse and preference learning shifts in LLMs without manual curation.
• Auditing Framing-Sensitive Behavioral Instability in Large Language Models for Mental Health Interactions — LLMs exhibit framing-sensitive behavioral instability in mental health contexts, rooted in internal representations.
• FLiP: Towards understanding and interpreting multimodal multilingual sentence embeddings — FLiP models effectively interpret sentence embeddings by linearly extracting lexical content, revealing underlying biases.
• Polar: A Benchmark for Evaluating Political Bias in LLMs — LLM political bias is context-dependent, varying by language, ideology, and geography.
• Cross-Lingual Consensus: Aligning Multilingual Cultural Knowledge via Multilingual Self-Consistency — LLMs can self-align cultural knowledge across languages by identifying consistent responses and transferring them to weaker language contexts.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.