Retrieval-Augmented Self-Taught Reasoning Model with Adaptive Chain-of-Thought for ASR Named Entity Correction

2026-02-16 · Source: cs.CL updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Natural Language Processing · Depth: Expert, long

Summary

A new retrieval-augmented self-taught reasoning model with adaptive chain-of-thought (RASTAR) has been developed to correct named entity errors in Automatic Speech Recognition (ASR) systems. This framework addresses the common issue of ASR misrecognizing domain-specific phrases, which can lead to significant downstream failures. RASTAR comprises two main components: a rephrasing language model (RLM) for robust named entity recognition (NER) that uses contextual semantic understanding, and an adaptive self-taught reasoning model (A-STAR) that dynamically adjusts its reasoning depth based on task difficulty. Experiments on the AISHELL-1 and Homophone datasets demonstrated RASTAR's effectiveness, achieving relative reductions in named entity character error rate (NE-CER) of 17.96% and 34.42%, respectively, compared to a DANCER baseline. The model also showed improved reasoning efficiency, particularly for larger models like Qwen3-8B, which reduced token usage by 30% and 21% on the respective datasets.

Key takeaway

For NLP engineers and research scientists working on ASR error correction, RASTAR offers a significant advancement in handling named entity misrecognitions. You should consider integrating its rephrasing language model for more robust NER and its adaptive chain-of-thought mechanism to improve correction accuracy while optimizing computational resources, especially for larger LLMs. This approach can lead to substantial reductions in named entity error rates, particularly in phonetically challenging scenarios.

Key insights

RASTAR improves ASR named entity correction by combining robust NER with adaptive, self-taught reasoning.

Principles

• Contextual semantic understanding enhances NER robustness.
• Adaptive reasoning depth reduces computational overhead.
• Self-training with preference optimization improves model performance.

Method

RASTAR uses a rephrasing language model for NER, followed by phonetic-level candidate retrieval. An adaptive self-taught reasoning model (A-STAR) then classifies problem difficulty to dynamically adjust chain-of-thought depth, trained via DPO on self-distilled data.

In practice

• Use RLM for NER in noisy ASR transcripts.
• Implement adaptive CoT to optimize LLM inference costs.
• Apply DPO for self-training reasoning models.

Topics

• Automatic Speech Recognition
• Named Entity Correction
• Retrieval-Augmented Generation
• Adaptive Chain-of-Thought
• Self-Taught Reasoning

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

Related on AIssential

• Ontology Memory-Augmented ASR Correction for Long Text-Speech Interleaved Conversations — Ontology memory significantly improves ASR correction in long, interleaved conversations by providing structured, dynamically updated contextual evidence.
• RASER: Recoverability-Aware Selective Escalation Router for Multi-Hop Question Answering — RASER reduces multi-hop QA token costs by selectively escalating retrieval based on one-shot RAG features.
• DiZiNER: Disagreement-guided Instruction Refinement via Pilot Annotation Simulation for Zero-shot Named Entity Recognition — Simulating human pilot annotation with LLMs to refine instructions significantly boosts zero-shot NER performance.
• RASC+: Retrieval-Constrained LLM Adjudication for Clinical Value Set Authoring — Retrieval-constrained LLM adjudication significantly improves clinical value set authoring by combining high-recall candidate generation with precise selection.
• Prompt Engineering for Named Entity Extraction from Portuguese Legal Documents — Prompt engineering with LLMs offers a viable alternative for legal NER in low-resource settings.
• Teaching Language Models to Check Grounded Claim Factuality with Human Test-Taking Strategies — Formulating factuality checking as a true/false reading comprehension task with test-taking strategies improves LLM efficiency and accuracy.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CL updates on arXiv.org.