University of Bucharest Boosts Romanian Coreference Resolution by 19 Points Without Training Data
Researchers from the University of Bucharest propose a cycle-consistent machine translation pipeline that bootstraps coreference resolution for languages with little to no training data.
Takeaways
- Core contribution: University of Bucharest researchers propose a cycle-consistent MT pipeline to build coreference resolution systems for low-resource languages with little or no training data.
- How it works: English CR data is translated to the target language via LLM, back-translated to English, and BERTScore similarity is used to weight each document's contribution to the training loss.
- Results: Tested on French, Hungarian, Romanian, and Russian -- Romanian (zero native data) jumps from 64.1 to 83.1 CoNLL F1; French improves from 79.7 to 85.1.
- New dataset: The team manually curated the first-ever Romanian coreference resolution test set, to be released publicly under the LDC User Agreement.
- Key finding: BERTScore outperforms BLEU as a cycle-consistency metric; the loss-weighting exponent p=3 is optimal and transfers across languages without retuning.
- Limitations: Relies on a commercial LLM (Claude Sonnet 4.6) for translation; potential for bias transfer; only four languages tested. Paper on arXiv.
Coreference resolution (CR) -- the NLP task of identifying all text spans that refer to the same real-world entity, like linking "she," "the CEO," and "Maria" across a document -- is one of those foundational capabilities that quietly powers machine translation, question answering, and document summarization. The catch: it works well almost exclusively in English. A team of researchers from the University of Bucharest has now published a paper proposing a practical pipeline to extend CR to low-resource languages, even ones with zero existing annotated data.
The dataset desert problem
The state of multilingual CR is stark. While the task is well-studied in English, comparatively less attention is dedicated to coreference resolution in other languages, especially low-resource ones. The CRAC 2025 shared task, a major annual benchmark for multilingual CR, covers 33 datasets spanning 19 languages -- but that still leaves hundreds of languages completely unaddressed. And zero-shot LLMs are not a reliable fallback: the CRAC 2025 results indicate that zero-shot LLMs lag far behind specialized models, with a clear gap of about 13% in terms of F1.
For Romanian specifically, the authors found no publicly available CR dataset at all -- making it a clean test case for a language that is entirely data-free at the start of the pipeline.
Translate, check the round trip, then train
The core idea is elegant: use English CR datasets (which are rich and well-annotated) as a source, translate them into the target language, and train a specialized model on the result. The tricky part is that machine translation is imperfect, and annotation spans -- the bracket markers that identify entity mentions -- can get dropped, misaligned, or garbled during translation. Bad translated samples will poison the model if treated equally.