The article investigates whether unsupervised dependency parsing can be evaluated in non-human species where gold standard annotations are unavailable. It argues that while this is generally considered impossible for human languages, it is feasible for non-human primates due to specific statistical properties of their communication.
- The feasibility relies on the fast decay of sequence length distribution found in primate vocalizations and gestures, which necessitates a high proportion of correct edges in the parsed tree.
- Human language sequences do not exhibit this fast decay property, making gold-standard-free evaluation significantly harder for humans than for other species.
- Recent advances in network science are applied to demonstrate that the structural constraints of non-human primate sequences allow for accuracy assessment without external reference data.