RspectNLP@LT-EDI 2026:Rubric-Driven Prompting for Safe Multilingual Counter Narrative Generation

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Natural Language Processing · Depth: Expert, quick

Summary

The RspectNLP@LT-EDI 2026 system introduces a zero-shot multilingual approach for generating safe counter-narratives against harmful online discourse targeting the LGBTQ+ community. This system, designed for English and Tamil, utilizes the google/flan-t5-base transformer model. It employs rubric-aligned prompts to guide response generation towards politeness, contextual relevance, and non-toxicity, operating without task-specific fine-tuning. Beam search decoding further controls the output. On English test data, the system achieved an overall score of 70.33% and an 81.82% contextual coherence score. However, performance significantly dropped on Tamil test data, scoring 33.57% overall with lower coherence and quality. These results highlight the effectiveness of structured prompting for English counter-narrative generation but also expose limitations of zero-shot multilingual models in low-resource language contexts.

Key takeaway

For NLP Engineers developing online safety systems, particularly for counter-narrative generation, you should consider rubric-driven prompting as an effective strategy for English. While this approach yields strong results (e.g., 70.33% overall score), be aware that zero-shot multilingual models like "flan-t5-base" may perform poorly in low-resource languages such as Tamil (33.57% overall). You might need language-specific fine-tuning or more robust data strategies for non-English deployments.

Key insights

Rubric-driven prompting with "flan-t5-base" generates safe counter-narratives effectively in English, but faces challenges in low-resource languages like Tamil.

Principles

• Structured prompting enhances safe generation.
• Zero-shot models struggle in low-resource languages.
• Counter-narratives offer constructive responses.

Method

A zero-shot multilingual system uses "google/flan-t5-base" with rubric-aligned prompts to guide politeness, contextual relevance, and non-toxicity. Beam search decoding controls response generation without fine-tuning.

In practice

• Generate empathetic counter-narratives.
• Apply rubric-aligned prompts for control.
• Use beam search for response quality.

Topics

• Counter-Narrative Generation
• Multilingual NLP
• Zero-Shot Learning
• Flan-T5
• Prompt Engineering
• Online Safety
• Low-Resource Languages

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

Related on AIssential

• Can AI Save India’s Dying Languages? | Building a Tulu LLM Without Training Data Ft. Prathamesh — Structured prompting and custom romanization enable LLMs to generate low-resource languages without extensive training data.
• RubiCap: Rubric-Guided Reinforcement Learning for Dense Image Captioning — RubiCap uses LLM-generated rubrics to provide fine-grained, structured reward signals for dense image captioning via reinforcement learning.
• Locale-Conditioned Few-Shot Prompting Mitigates Demonstration Regurgitation in On-Device PII Substitution with Small Language Models — Locale-conditioned few-shot prompting prevents SLMs from regurgitating PII substitution demonstrations.
• Automatic Generation of Highlights for Academic Paper Via Prompt-based Learning — Prompt-based learning with LLMs can generate academic paper highlights effectively without extensive labeled data.
• Feedback-to-Rubrics: Can We Learn Expert Criteria from Inline Comments? — Inline comments can be distilled into reusable natural-language rubrics, enhancing LLM-driven writing and review support.
• PromptMN: Pseudo Prompting Language — PromptMN structures natural language prompts with typed directives, enhancing clarity and reliability in human-AI interaction.

Open in AIssential →

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