Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

DACE For Railway Acronym Disambiguation

El Mokhtar Hribach, Oussama Mechhour, Mohammed Elmonstaser, Yassine El Boudouri, Othmane Kabal

TL;DR

The paper tackles Acronym Disambiguation in French railway documentation, a domain with high terminology ambiguity and limited annotated data. It introduces DACE, a modular framework that fuses Dynamic Prompting, Retrieval Augmented Generation, Contextual Selection, and Ensemble Aggregation to improve LLM-driven disambiguation while curbing hallucinations. On the TextMine'26 dataset, DACE achieves a top private F1 of 0.9069 and demonstrates stable performance across splits, underscoring its robustness to low-resource and domain-shift conditions. Ablation studies confirm that both adaptive prompting and knowledge grounding substantially contribute to performance, with ensemble reasoning further enhancing reliability. The framework’s general design suggests applicability beyond railways, including multilingual settings and graph-based knowledge integration future work.

Abstract

Acronym Disambiguation (AD) is a fundamental challenge in technical text processing, particularly in specialized sectors where high ambiguity complicates automated analysis. This paper addresses AD within the context of the TextMine'26 competition on French railway documentation. We present DACE (Dynamic Prompting, Retrieval Augmented Generation, Contextual Selection, and Ensemble Aggregation), a framework that enhances Large Language Models through adaptive in-context learning and external domain knowledge injection. By dynamically tailoring prompts to acronym ambiguity and aggregating ensemble predictions, DACE mitigates hallucination and effectively handles low-resource scenarios. Our approach secured the top rank in the competition with an F1 score of 0.9069.

DACE For Railway Acronym Disambiguation

TL;DR

The paper tackles Acronym Disambiguation in French railway documentation, a domain with high terminology ambiguity and limited annotated data. It introduces DACE, a modular framework that fuses Dynamic Prompting, Retrieval Augmented Generation, Contextual Selection, and Ensemble Aggregation to improve LLM-driven disambiguation while curbing hallucinations. On the TextMine'26 dataset, DACE achieves a top private F1 of 0.9069 and demonstrates stable performance across splits, underscoring its robustness to low-resource and domain-shift conditions. Ablation studies confirm that both adaptive prompting and knowledge grounding substantially contribute to performance, with ensemble reasoning further enhancing reliability. The framework’s general design suggests applicability beyond railways, including multilingual settings and graph-based knowledge integration future work.

Abstract

Acronym Disambiguation (AD) is a fundamental challenge in technical text processing, particularly in specialized sectors where high ambiguity complicates automated analysis. This paper addresses AD within the context of the TextMine'26 competition on French railway documentation. We present DACE (Dynamic Prompting, Retrieval Augmented Generation, Contextual Selection, and Ensemble Aggregation), a framework that enhances Large Language Models through adaptive in-context learning and external domain knowledge injection. By dynamically tailoring prompts to acronym ambiguity and aggregating ensemble predictions, DACE mitigates hallucination and effectively handles low-resource scenarios. Our approach secured the top rank in the competition with an F1 score of 0.9069.

Paper Structure

This paper contains 20 sections, 3 equations, 3 figures, 2 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Acronym Disambiguation
  4. In-Context Learning and Prompt Engineering
  5. Proposed Approach
  6. Problem Formulation
  7. The DACE Framework Overview
  8. Dynamic Prompting
  9. Contextual Selection
  10. Retrieval Augmented Generation (Knowledge Grounding)
  11. Ensemble Aggregation
  12. 1. Subset Selection:
  13. 2. Majority Voting with Tie-Breaking:
  14. 3. Competence Fallback:
  15. Experiments
  16. ...and 5 more sections

Figures (3)

  • Figure 1: Architecture of the DACE framework.
  • Figure 2: Distribution of unique acronyms across the training and test sets.
  • Figure 3: Analysis of candidate and correct expansion distributions in the training set. Figure \ref{['fig_options_chart']} shows the variability in the number of candidate options, and Figure \ref{['fig_correct_options_chart']} shows the distribution of correct answers.