Cross-Lingual Evidence-Based Strategies for Identifying Fabrications in Neural Translation Systems

Abstract

This paper investigates cross-lingual evidence-based strategies for uncovering fabrications that emerge in neural translation systems, focusing on the phenomenon known as hallucination in machine-generated texts. Although neural architectures have achieved remarkable performance in many translation tasks, they remain prone to generating content that is factually ungrounded or entirely fabricated. Such fabrications compromise reliability, especially in domains where accuracy is mandatory. The proposed framework integrates linguistic alignment, bilingual term matching, and extrinsic verification checks across multiple language pairs. Emphasis is placed on constructing domain-specific corpora that underscore knowledge-intensive expressions, enabling robust identification of fabricated segments. Empirical analysis covers a spectrum of neural machine translation models, investigating how each architecture handles rare terms and nuanced syntactic constructs. The paper also develops a robust suite of metrics designed to quantify hallucination severity, leveraging lexical similarity measures and cross-entropy differentials. Results demonstrate that employing external knowledge sources and semantic aligners can reduce fabrication rates across a variety of languages, thereby enhancing translation integrity. The implications of this research extend to areas such as cross-border communications, international legal proceedings, and medical translations. By synthesizing empirical findings, this study offers a nuanced roadmap for future explorations in cross-lingual integrity verification, highlighting the evolving interplay between data-driven models and linguistic fidelity.