A deep learning framework for translation quality estimation using dual stream cross lingual attention and gated mixture of experts prediction

Abstract

Translation quality estimation (TQE)—predicting translation quality without reference translations—is increasingly important in natural language processing. Current neural TQE methods encode source and translated texts in a single concatenated representation, limiting fine-grained cross-lingual alignment, and employ homogeneous regression heads that ignore translation error heterogeneity. This paper proposes DL-TQE, a deep learning architecture comprising two components: (1) a Dual-Stream Cross-Lingual Attention Network (DSCAN) with parallel Transformer encoders connected by explicit cross-lingual attention bridges at multiple layers, and (2) a Gated Mixture-of-Experts Quality Predictor (GMoE-QP) that routes quality-relevant features through specialised expert sub-networks conditioned on translation complexity signals. Evaluated on WMT 2020–2023 Quality Estimation benchmarks across four English-centric language pairs (En–De, En–Zh, En–Cs, En–Ja), DL-TQE achieves a Pearson correlation of 0.641 ± 0.003 on the En–De benchmark, outperforming TransQuest (0.612) and CometKiwi (0.630) by 2.9 and 1.1% points respectively (p < 0.05, Williams test). Ablation analysis confirms that DSCAN and GMoE-QP independently contribute 2.3 and 1.8%-point improvements, respectively. Expert routing analysis reveals domain-sensitive specialisation, and error-type disaggregated evaluation demonstrates consistent advantages across lexical, morpho-syntactic, and semantic error categories within the tested WMT benchmark settings. These findings suggest that explicit cross-lingual alignment modelling and error-type-sensitive prediction offer a promising direction for sentence-level TQE on English-centric language pairs.