Although existing Chinese word segmentation systems have achieved substantial progress on standard benchmarks, prediction disagreements among heterogeneous models remain prevalent when processing texts containing complex ambiguities and out-of-vocabulary words, and traditional static ensemble methods such as majority voting often fail to make reliable decisions in low-consensus scenarios. To address this issue, this paper proposes SegFusion, a stacked heterogeneous ensemble framework for Chinese word segmentation based on word lattice re-scoring. The framework first constructs a candidate word lattice to consolidate diverse outputs from heterogeneous segmenters into a unified lattice representation, and then incorporates unsupervised statistical features, including mutual information and branching entropy, as external discriminative evidence to perform dynamic arbitration at the word level, followed by global decoding to obtain the optimal segmentation path. Experimental results on multiple standard datasets demonstrate that SegFusion consistently outperforms individual models and mainstream ensemble baselines in terms of overall segmentation performance and out-of-vocabulary (OOV) recall. In particular, on the MSR dataset with severe ambiguity, SegFusion achieves improvements of 3.71% in F1 score and 4.10% in OOV recall. Further fine-grained analysis shows that the introduction of unsupervised statistical features effectively mitigates model consistency bias in low-support scenarios. These results indicate that integrating language statistical priors independent of training data into the ensemble arbitration stage is an effective way to enhance the robustness and consistency of Chinese word segmentation systems.
Wen et al. (2026) studied this question.