Sequence-Smoothened Semi-Supervised POS Tagging for Assamese Language Using Cluster-Anchored Weak Supervision and HMM Viterbi Decoding

This paper presented a comprehensive study of semi-supervised part-of-speech tagging for Assamese, a low-resource language characterized by limited annotated data and high lexical ambiguity. Through a sequence of progressively structured experiments, we investigated the effectiveness of self-training, graph-based label propagation, cluster-anchored weak supervision, and sequence-level decoding using a Hidden Markov Model with Viterbi inference. Our findings demonstrate that naïve self-training with confi-dence filtering is ineffective in this setting. Despite high poste-rior confidence, pseudo-labels generated from a weakly trained seed model reinforce early errors, leading to performance degradation across iterations. This confirms that confidence alone is not a reliable proxy for correctness in low-resource POS tagging and highlights the need for stronger inductive bias. Graph-based label propagation provides such an inductive bias by exploiting distributional similarity among word types. While this approach significantly outperforms random and majority baselines, its reliance on word-type representations limits contextual expressiveness. As a result, frequent nouns and verbs benefit most, whereas modifiers and function words remain challenging due to over-smoothing and lack of contex-tual differentiation. The core contribution of this work is the cluster-anchored weak supervision framework, which bridges the gap be-tween unsupervised induction and supervised POS tagging. By transforming high-purity distributional clusters into la-beling functions, the approach generates large-scale, high-confidence weak supervision without manual annotation. This method achieves near-complete token coverage and yields substantial improvements in both accuracy and weighted F1-score, demonstrating that distributional clusters encode strong syntactic regularities. Building upon this foundation, we further introduced se-quence smoothing using HMM-based Viterbi decoding to explicitly model sentence-level structure. While token-level classifiers already produce strong local predictions, Viterbi decoding enforces global consistency by incorporating learned transition preferences. Empirically, this results in syntactically coherent tag sequences, particularly benefiting nouns, verbs, and conjunctions. Although a small reduction in token-level accuracy is observed, the overall model achieves a better balance between local evidence and global structure, which is essential for linguistically plausible tagging. Overall, this work shows that effective semi-supervised POS tagging in low-resource languages requires a combination of distributional inductive bias and sequence-level modeling. The proposed framework avoids the pitfalls of self-training, scales without additional annotation, and produces robust results across major POS categories. Beyond Assamese, the methodology is language-agnostic and readily applicable to other low-resource and morphologically rich languages. Future work may explore neural sequence models initial-ized with cluster-based weak supervision, adaptive transition modeling, and joint learning of clustering and tagging. We believe that the proposed approach provides a principled and practical direction for advancing syntactic analysis in low-resource settings.

[1]     L. R. Rabiner, “A tutorial on hidden markov models and selected applications in speech recognition,” Proceedings of the IEEE, vol. 77, no. 2, pp. 257–286, 1989. [2]     A. J. Viterbi, “Error bounds for convolutional codes and an asymptoti-cally optimum decoding algorithm,” IEEE Transactions on Information Theory, vol. 13, no. 2, pp. 260–269, 1967. [3]     T. Brants, “Tnt – a statistical part-of-speech tagger,” in Sixth Applied Natural Language Processing Conference. Association for Computational Linguistics, 2000, pp. 224–231. [Online]. Available: https://aclanthology.org/A00-1031/ [4]     J. D. Lafferty, A. McCallum, and F. C. N. Pereira, “Conditional random fields: Probabilistic models for segmenting and labeling sequence data,” in Proceedings of the Eighteenth International Conference on Machine Learning (ICML), 2001, pp. 282–289. [Online]. Available: https://dl.acm.org/doi/10.5555/645530.655813 [5]     P. Brown et al., “Class-based n-gram models of natural language,” Computational Linguistics, 1992. [6]     C. Christodoulopoulos, S. Goldwater, and M. Steedman, “Two decades of unsupervised pos induction,” Proceedings of EMNLP, 2010. [7]     D. Yarowsky, “Unsupervised word sense disambiguation rivaling supervised methods,” in Proceedings of the 33rd Annual Meeting of the Association for Computational Linguistics (ACL), 1995, pp. 189–196. [Online]. Available: https://aclanthology.org/P95-1026/ [8]     A. Ratner et al., “Snorkel: Rapid training data creation with weak supervision,” VLDB, 2017. [9]     K. Kann et al., “Weakly supervised pos taggers perform poorly on truly low-resource languages,” in Proceedings of the AAAI Conference on Artificial Intelligence, 2020. [Online]. Available: https://ojs.aaai.org/index.php/AAAI/article/view/6317 [10]    D. Das and S. Petrov, “Unsupervised part-of-speech tagging with bilingual graph-based projections,” in Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies. Association for Computational Linguistics, 2011, pp. 600–609. [Online]. Available: https://aclanthology.org/ P11-1061/