TITLE:
AMorph: An End-to-End Morpheme Level Natural Language Processing Pipeline for Amharic
AUTHORS:
Natnael Tamirat Molla, Shi Ming
KEYWORDS:
Amharic, Morphological Segmentation, Finite-State Morphology, Weak Supervision, Connectionist Temporal Classification (CTC), Multi-Task Learning, Dependency Parsing, Universal Dependencies
JOURNAL NAME:
Open Access Library Journal,
Vol.13 No.2,
February
12,
2026
ABSTRACT: Morphological segmentation is foundational for Natural Language Processing in morphologically rich languages, such as Amharic, yet progress is constrained by limited gold annotations and fragmented toolchains. We present an end-to-end framework that jointly addresses data creation and modeling for Amharic segmentation, part-of-speech tagging, and dependency parsing. Our approach begins with a silver-data generation pipeline that bootstraps segmentation labels from a rule-based analyzer and refines them through automated review, large language model assisted verification, and human-in-the-loop correction. Using the resulting supervision, we train an XLM-RoBERTa segmenter with a connectionist temporal classification (CTC) based character transduction objective, enabling reliable morpheme boundary prediction without explicit character-morpheme alignment. We then introduce a unified multi-task toolkit model that replaces the common practice of training separate systems per task. A shared pretrained encoder is jointly optimized for POS tagging and dependency parsing to better capture cross-task linguistic regularities while remaining parameter-efficient. The observed morpheme segmentation, POS tagging, and dependency parsing results support the conclusion that analyzer-bootstrapped supervision combined with multilingual pretrained encoders is effective for Amharic morphosyntactic modeling in low-resource settings. We release an open-source toolkit with simple APIs and an interactive visualization interface, enabling users to run the pipeline and inspect intermediate and final outputs for practical Amharic NLP development. API, documentation, and pre-trained models are available at https://github.com/Netela-lab/AMorph.