A Multi-way Parallel Named Entity Annotated Corpus for English, Tamil and Sinhala

TL;DR

This work introduces a novel multi-way parallel English–Sinhala–Tamil named entity annotated corpus and demonstrates that pre-trained multilingual language models, particularly multilingual variants of XLM-R, achieve strong NER performance on Sinhala and Tamil. It analyzes language-specific versus multilingual fine-tuning, and shows that a single multilingual model can effectively handle all three languages. The dataset enables robust cross-language transfer and improves downstream NMT when NE information is leveraged, as evidenced by a significant BLEU gain in a case study. Public release of the data facilitates further research in NER for low-resource languages and improves MAL (multilingual) NLP tooling for Sri Lankan and Indian language contexts.

Abstract

This paper presents a multi-way parallel English-Tamil-Sinhala corpus annotated with Named Entities (NEs), where Sinhala and Tamil are low-resource languages. Using pre-trained multilingual Language Models (mLMs), we establish new benchmark Named Entity Recognition (NER) results on this dataset for Sinhala and Tamil. We also carry out a detailed investigation on the NER capabilities of different types of mLMs. Finally, we demonstrate the utility of our NER system on a low-resource Neural Machine Translation (NMT) task. Our dataset is publicly released: https://github.com/suralk/multiNER.

Figures (3)

• Figure 1: Sample English/Tamil/Sinhala sentences annotated with CONLL03 tag set, following the BIO format.
• Figure 2: Architecture of the Bi-LSTM CRF network with affix features yadav2018deep
• Figure 3: DEEPhu2021deep Architecture