Year: 2026
Type: conference-paper
Abstract: We present a systematic study of paraphrase detection in Tamil through the construction and evaluation of a unified benchmark dataset derived from three widely used English paraphrase corpora: QQP, PAWS, and MRPC.The dataset is generated using a translation and verification pipeline that combines semantic similarity analysis, round-trip consistency checks, classifier agreement evaluation, and human assessment to improve translation reliability and semantic preservation.Using this benchmark, we evaluate five multilingual transformer models (mBERT, XLM-R, IndicBERT, MuRIL, and Distilm-BERT) alongside a compact Tamil-specific encoder, TLMR, pretrained on 525M Tamil tokens.We further examine the representational quality of these models through embeddingbased classification using lightweight machine learning classifiers, including Logistic Regression, SVM, and XGBoost.In addition, we introduce a Tamil-oriented efficiency evaluation framework that analyzes tradeoffs between masked language modeling performance, vocabulary utilization, and Tamil script fidelity.Experimental results show that multilingual models such as MuRIL achieve strong downstream paraphrase detection performance, while Tamil-specific pretraining in TLMR improves efficiency-oriented language modeling characteristics.Our findings provide a benchmark and analysis framework for future research on semantic understanding in Tamil and other low-resource languages.
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7166807973
FWCI:
Cites:
Cited by:
Related to: