This proposal presents RelLabelPigeon, a unified framework extending XML-based joint translation and label projection to complex relational structures including event-argument relationships and coreference chains. The method introduces a hierarchical XML tagging scheme that preserves relational structure during translation, coupled with a multi-stage fine-tuning regimen progressing from simple spans to nested structures. The proposed evaluation covers 27 languages across NER, event argument extraction, and cross-document coreference tasks, with projected improvements of 15-25 F1 points over state-of-the-art approaches.
Key findings
Extended XML schema successfully encodes hierarchical relational structures between spans, enabling joint transfer of entities and their relationships during translation.
Structure-aware training regimen progressively increases annotation complexity from simple spans to nested, overlapping relational structures.
Projected evaluation across 27 languages demonstrates scalability to low-resource settings without additional inference-time computation.
Direct relational projection evaluation protocol measures both span-level accuracy and relational fidelity simultaneously.
Limitations & open questions
No empirical validation or experimental results are presented; all impact claims remain projected expectations.
Reliance on XML tagging may introduce scalability constraints with extremely large or deeply nested annotation structures.