Error Diversity Matters: An Error-Resistant Ensemble Method for Unsupervised Dependency Parsing

TL;DR

This work tackles unsupervised dependency parsing by assembling diverse, unsupervised parsers and combining their outputs through a post hoc, diversity-aware ensemble framework. It introduces a minimum Bayes risk–based aggregation using the unlabeled attachment score as the similarity metric and a forward-selected ensemble that explicitly optimizes both average performance and error diversity via a novel society entropy measure. Empirical results on the WSJ Penn Treebank show the approach outperforming each individual parser and prior ensemble methods, achieving state-of-the-art unsupervised parsing performance and demonstrated robustness to weak components. The framework offers practical advantages for low-resource languages and domains by reducing the impact of correlated errors and enabling efficient, scalable ensemble construction. Mathematical formulations include $A^*$ optimization under $UAS$ similarities and a $\mathcal{O}(n^3)$ dependency-DP solution, and the diversity objective employs $\operatorname{SE}$ based on per-word head distributions $\operatorname{SD}$.$

Abstract

We address unsupervised dependency parsing by building an ensemble of diverse existing models through post hoc aggregation of their output dependency parse structures. We observe that these ensembles often suffer from low robustness against weak ensemble components due to error accumulation. To tackle this problem, we propose an efficient ensemble-selection approach that considers error diversity and avoids error accumulation. Results demonstrate that our approach outperforms each individual model as well as previous ensemble techniques. Additionally, our experiments show that the proposed ensemble-selection method significantly enhances the performance and robustness of our ensemble, surpassing previously proposed strategies, which have not accounted for error diversity.

Figures (4)

• Figure 1: (a) Performances of individuals. (b) Performances of ensembles by adding individuals in (a) from left to right. Numbers are UASs on the WSJ test set, reported by the mean and the standard deviation through five runs. The experiment setup is detailed in §\ref{['sec:settings']}.
• Figure 2: Ensemble performance by the number of selected ensemble components using different selection strategies. Ensemble selection happens over 25 individuals (five runs for each of CRFAE, NDMV, NE-DMV, L-NDMV, and Sib-NDMV). Results are split into two figures for easier reading.
• Figure 3: UAS by dependents' POS tags on the WSJ test set. The ten represented POS tags cover around 78% of the cases.
• Figure 4: A dependency parse structure represented by directed dependency arcs (a) and DPST (b).