International Conference on Computer and Knowledge Engineering

Home / 15th International Conference on Computer and Knowledge Engineering

ParsHomo: A T5-Powered Approach to High-Precision Persian Homograph Disambiguation

Authors :

Hasan Jalali¹ Taha Mohaddesi²

1- School of Electrical and Computer Engineering College of Engineering, University of Tehran 2- ASMIN AI Research Center

Keywords :

Homograph Disambiguation،Persian Language Processing،Grapheme-to-Phoneme Conversion،T5 Transformer،Natural Language Processing،Text-to-Speech

Abstract :

This paper presents ParsHomo, a high-precision model for Persian homograph disambiguation using fine-tuned T5-based architectures (mT5 and ByT5). Starting from the HomoRich dataset, we manually curated and refined the human-labeled portion to ensure exceptional data quality. The cleaned dataset was used to fine-tune T5 models, leading to substantial improvements over previous systems. We evaluated various input formatting strategies—including question-style, label-based, and separator-based approaches—and used the trained model to relabel large-scale datasets such as HomoRich-GPT4o and GE2PE, resulting in a bootstrapped corpus of over 4.19 million sentences. ParsHomo offers the most accurate and scalable homograph disambiguation system for Persian to date, demonstrating the effectiveness of transformer-based models for contextaware phoneme generation in low-resource languages. Our bestperforming model (mT5 with label-based format) achieved an accuracy of 94.55% and a Character Error Rate of 1.71%, outperforming prior approaches by a wide margin.