Small Language Models Show Limits in Cross-Lingual Sentiment Detection

A study published in Neural Computing and Applications evaluated instruction-tuned Small Language Models in Target Aspect Sentiment Detection. These settings remain difficult due to semantic alignment gaps between different linguistic systems. The study presents a comprehensive benchmark evaluating these models across eight diverse languages using Parameter-Efficient Fine-Tuning (PEFT).

The investigation focused on three distinct prompt paradigms to determine which approach best facilitates cross-lingual alignment. The researchers observed that Structured Extraction consistently yielded superior performance in aligning semantics across languages.

Architectural differences between model types also emerged as a significant factor in performance. The study found that decoder-only SLMs generally led in semantic alignment across most target languages. However, encoder-decoder architectures demonstrated unique resilience in morphologically complex, low-resource languages.

Empirical evidence from the benchmark suggests that the transfer mechanisms at play correlate more strongly with deep semantic abstraction than with superficial subword overlap.

The researchers identified a substantial instruction tuning gap. Furthermore, systemic issues were observed in resolving exact aspect boundaries and handling complex morphology. These errors point to capacity ceilings in the evaluated SLMs.

The study was published on 12 November 2026 in Volume 38, article number 882.

Subject of Research: Cross-lingual Target Aspect Sentiment Detection using instruction-tuned Small Language Models

Article Title: Enhancing cross-lingual Target Aspect Sentiment Detection via instruction-tuned Small Language Models

Article References: Phuc, N. X., Van Thin, D., Ngoc Hao, D., & Nguyen, N. L.-T. (2026). Enhancing cross-lingual Target Aspect Sentiment Detection via instruction-tuned Small Language Models. Neural Computing and Applications, 38(21), Article 882. https://doi.org/10.1007/s00521-026-12615-w

Image Credits: AI Generated

DOI: 10.1007/s00521-026-12615-w

Keywords: Target Aspect Sentiment Detection, Small Language Models, Cross-lingual NLP, Parameter-Efficient Fine-Tuning, Instruction Tuning, Zero-shot Learning, Joint-shot Learning, Semantic Alignment, Structured Extraction, Decoder-only Models, Encoder-decoder Models, Low-resource Languages, Morphological Complexity, Subword Overlap, Neural Computing and Applications, Multilingual Benchmark, Aspect Boundary Resolution, Instruction Tuning Gap, Language Model Capacity, Vietnamese Research Team