Lispector: Fine-tuning of Language Models for Grammar and Spelling Correction in Brazilian Portuguese

2026-04-12 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, quick

Summary

Lispector is a new family of specialized language models designed for grammar and spelling correction in Brazilian Portuguese. This work compares two inference strategies: supervised fine-tuning and few-shot prompting with larger models. Utilizing a dataset of 4,500 real user text pairs, corrected by linguists, two Lispector variants with different parameter sizes were evaluated. The models were assessed using BLEU, GLEU, METEOR, and ROUGE metrics. Results show that smaller, supervised fine-tuned models consistently outperform larger models using only prompting, with Lispector small achieving notable gains in GLEU (+12%) and BLEU (+13%). These fine-tuned models also exhibit more predictable and conservative behavior, alongside lower latency, making them suitable for industrial assisted writing applications.

Key takeaway

For AI Engineers developing grammar and spelling correction tools for Brazilian Portuguese, fine-tuning smaller language models like Lispector offers superior performance and computational efficiency compared to few-shot prompting larger models. You should consider supervised fine-tuning to achieve more predictable model behavior and lower latency, which are critical for industrial applications.

Key insights

Supervised fine-tuning of smaller models outperforms few-shot prompting for grammar correction in Brazilian Portuguese.

Principles

• Fine-tuning yields predictable, conservative model behavior.
• Smaller, fine-tuned models can surpass larger, prompted models.

Method

The method involves supervised fine-tuning of language models using a linguist-corrected dataset of real user texts, followed by evaluation with BLEU, GLEU, METEOR, and ROUGE metrics.

In practice

• Use fine-tuning for specific language tasks.
• Prioritize smaller models for efficiency.
• Employ linguist-corrected datasets for training.

Topics

• Lispector
• Language Model Fine-tuning
• Brazilian Portuguese
• Grammar and Spelling Correction
• Few-shot Prompting

Best for: AI Engineer, Research Scientist, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• Efficient Fine-Tuning Methods for Portuguese Question Answering: A Comparative Study of PEFT on BERTimbau and Exploratory Evaluation of Generative LLMs — PEFT methods efficiently fine-tune encoder-based models for low-resource languages, outperforming generative LLMs in resource usage.
• CURUPIRA: Clever guard for harm and linguistic prompt mitigation in Brazilian Portuguese — Targeted fine-tuning improves LLM safety classification for underrepresented languages like Brazilian Portuguese.
• GIFT: Guided Fine-Tuning and Transfer for Enhancing Instruction-Tuned Language Models — GIFT enhances instruction-tuned models by guiding base model adapter training with instruction model confidence signals.
• Think in Sentences: Explicit Sentence Boundaries Enhance Language Model's Capabilities — Explicitly modeling sentence boundaries with delimiters enhances LLM reasoning by inducing structured, sentence-by-sentence processing.
• TLPO: Token-Level Policy Optimization for Mitigating Language Confusion in Large Language Models — TLPO uses token-level policy optimization to mitigate language confusion in LLMs without degrading general capabilities.
• System Report for CCL25-Eval Task 5: New Dataset and LoRA-Fine-Tuned Qwen2.5 — Domain-specific datasets and LoRA fine-tuning significantly enhance LLM performance in specialized tasks like classical poetry analysis.

Open in AIssential →

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