On the feasibility of dependency parsing of non-human sequences without a gold standard. Is evaluation possible in other species?

TL;DR

The paper shows that unsupervised dependency parsing is feasible for non-human primate communication sequences without a gold standard, due to the rapid decay in sequence length distribution. This property enables high accuracy in edge recovery even without ground truth, unlike in human language where such decay is absent. Evaluation without a gold standard becomes possible in non-human species but remains challenging in human language.

Why It Matters

It opens the door to evaluating linguistic structure in non-human animal communication using computational methods, enabling comparative studies of syntactic complexity across species. This advances our understanding of the evolution of structured communication and provides tools for analyzing natural behavioral sequences without relying on human-annotated data.

Key Findings

• Non-human primate vocalization sequences exhibit a fast decay in sequence length distribution, which constrains possible dependency structures.
• This structural property implies that unsupervised parsers can recover a high proportion of correct dependency edges even without a gold standard.
• In contrast, human language sequences lack this decay, making evaluation without ground truth significantly harder and less reliable.