Consensus-Based Multi-Source Cross-Lingual Label Projection for Persian Named Entity Recognition

Consensus-Based Multi-Source Cross-Lingual Label Projection for Persian Named Entity Recognition

Reyhaneh Boroomand1 Marziea Rahimi2

1) Faculty of Computer Engineering, Shahrood University of Technology, Shahrood, Iran.
2) Faculty of Computer Engineering, Shahrood University of Technology, Shahrood, Iran.

IOI Code : AAHD-ABKDHZ
Publication : National Conference on Information Technology, Nanotechnology, Artificial Intelligence and Technological Futures Studies(itnaf.ir)
Abstract :
    The absence of large-scale accurately annotated corpora still remains a bottleneck for NER in low-resource languages. Cross-lingual label projection is an inexpensive alternative to manual annotation, but traditional single-source transfer pipelines are very sensitive to translation errors, word alignment failures, and inconsistent entity boundaries. We propose a consensus-based multi-source cross-lingual label projection framework to build a high-quality Persian NER dataset in this paper. Traditional approaches transfer annotations from one high-resource language. Conversely, the proposed framework exploits multiple source languages: English, German, and French, to generate complementary label projections into Persian. Each source stream undergoes source-side NER annotation, neural machine translation, word-to-word alignment, span-level alignment, and preliminary BIO label projection. Then, a confidence-aware consensus module integrates the projected annotations by jointly considering the source reliability, the lexical and contextual alignment confidence, the span overlap, and the agreement on the entity type. The framework also further applies boundary reconciliation and BIO-consistency validation to resolve conflicting projections and remove low-confidence annotations. The dataset is evaluated by training representative sequence-labeling and multilingual Transformer models such as BiLSTM-CRF, mBERT, and XLM-RoBERTa. The experiments are designed to evaluate the contribution of each source language and each refinement component by comparative and ablation analyses. The approach presented suggests a scalable and language-independent approach to reduce annotation noise and improve the reliability of automatically constructed NER resources for Persian and other low-resource languages.

Keywords : named entity recognition cross lingual transfer multi source learning label projection persian nlp low resource languages data construction