Scholarlink

<p>Large language models are remarkably fluent but often produce unfaithful summaries. Fabricated or omitted entities are particularly harmful in low–resource languages such as Hindi. We propose a reinforcement-learning approach that fine-tunes multilingual models with a reward based on the <i>Entity Hallucination Index</i> (EHI). Using a Hindi named-entity recogniser to identify entities in the source and summary, the EHI reward encourages grounded entity mentions while penalising hallucinations. Our training pipeline performs preprocessing, attaches LoRA adapters for efficient fine-tuning, and evaluates with entity metrics (EHI, EF1), ROUGE and a multilingual semantic judge. Experiments on the XL-Sum Hindi corpus with two base models (Mistral and LLaMA-2) show consistent, albeit modest, EHI improvements with negligible impact on ROUGE and small EF1 trade-offs. Statistical tests indicate that some gains are significant. This demonstrates the feasibility of entity-focused rewards for improving factual consistency in low-resource summarisation. Future work will enhance Hindi NER reliability and extend the reward to relation-level hallucination.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Reward-Driven Fine-Tuning with Entity Hallucination Index for Low-Resource Hindi Abstractive Summarisation

  • 期刊论文

  • 发表日期：

    2026年1月13日

查看全文
  • Praveenkumar Katwe,
  • Rakesh Chandra Balabantaray,
  • Kali Prasad Vittala

摘要

Large language models are remarkably fluent but often produce unfaithful summaries. Fabricated or omitted entities are particularly harmful in low–resource languages such as Hindi. We propose a reinforcement-learning approach that fine-tunes multilingual models with a reward based on the Entity Hallucination Index (EHI). Using a Hindi named-entity recogniser to identify entities in the source and summary, the EHI reward encourages grounded entity mentions while penalising hallucinations. Our training pipeline performs preprocessing, attaches LoRA adapters for efficient fine-tuning, and evaluates with entity metrics (EHI, EF1), ROUGE and a multilingual semantic judge. Experiments on the XL-Sum Hindi corpus with two base models (Mistral and LLaMA-2) show consistent, albeit modest, EHI improvements with negligible impact on ROUGE and small EF1 trade-offs. Statistical tests indicate that some gains are significant. This demonstrates the feasibility of entity-focused rewards for improving factual consistency in low-resource summarisation. Future work will enhance Hindi NER reliability and extend the reward to relation-level hallucination.