JCT 2026 - SemEval Task 5

AbstractThe system integrates a generative Large Language Model (Llama-3 8B, fine-tuned via LoRA) with a dual-expert bidirectional cross-encoder (DeBERTa-v3-large) optimized for both semantic similarity and Natural Language Inference (NLI). By aggregating these complementary models, the system effectively captures complex contextual dependencies. In the official test set, our architecture ranked 22nd out of 79 systems, achieving a Spearman Rank Correlation of 0.71 and an accuracy within the standard deviation of 82.04%.