Determinative Detection in Egyptian Hieroglyphic Sequences: A Sequence-Labelling Benchmark and Multi-Agent LLM Baseline

Abstract Determinatives, silent classifier signs that mark semantic fields and signal word boundaries, are a central feature of the ancient Egyptian writing system, and their automatic detection is a prerequisite for lemmatization, morphological parsing, and translation. Yet the task has received little dedicated attention as a language-processing problem, and no shared benchmark exists. This paper makes three contributions. First, we formulate determinative detection in sequences encoded in Manuel de Codage notation as a sequence-labelling task and release an expert-annotated benchmark of 99 hieroglyphic sequences, structured around a knowledge base of determinative signs organised into semantic categories following the Gardiner Sign List. Second, we propose a multi-agent method combining large language model (LLM) reasoning with deterministic domain constraints: two independent LLM agents analyse each sequence from semantic and structural perspectives, and a deterministic scoring module aggregates their evidence through a five-tier sign categorisation and context-dependent thresholds. Third, we evaluate the method under a nested cross-validation protocol in which every sequence serves as a held-out test case exactly once, with inner Bayesian optimisation of the scoring parameters. The method attains a macro-averaged F1 of 81.94% ± 2.19% and a micro-averaged F1 of 87.50% ± 2.50%, the low cross-fold variance quantifying its robustness. Beyond a single prediction, it produces ranked alternative interpretations with explicit promotion conditions and natural-language justifications, making it a transparent, collaborative resource for computational Egyptology, a low-resourced ancient language, rather than a black-box classifier.