Construction of urdu paraphrase types corpus (UPTC‑23) to detect fourteen urdu paraphrase types using multilingual BERT and distilBERT
Year: 2026
Type: article
Abstract: This study addresses the limited availability of linguistic resources for fine-grained Urdu paraphrase types identification by developing a manually annotated corpus and evaluating multilingual transformer models for a fourteen-class sentence-level classification task. A new Urdu Paraphrase Types Corpus (UPTC-23) was constructed comprising five thousand sentence pairs, manually annotated into fourteen Urdu paraphrase types based on the linguistically motivated typology. Sentence pairs compiled from the COUNTER corpus, online Urdu news articles, and manually curated linguistic examples to ensure comprehensive coverage of all fourteen Urdu paraphrase types. The resulting sentence pairs are subsequently subjected to manual preprocessing, quality verification, and annotation using predefined annotation guidelines. To evaluate the proposed corpus, multilingual BERT and DistilBERT models fine-tuned on the annotated dataset were used to identify the Urdu paraphrase type of each Urdu sentence pair by classifying it into one of the fourteen predefined Urdu paraphrase types. Experimental results demonstrate that the multilingual BERT-base-cased achieved the best performance, obtaining 92.97% accuracy and a macro-F1 score of 92%, while the multilingual DistilBERT-base-cased model achieved 91.87% accuracy and a macro-F1 score of 92%. Unlike existing Urdu paraphrase resources, which primarily support paraphrase identification or semantic textual similarity tasks, UPTC-23 provides fine-grained annotations of fourteen Urdu paraphrase types to support automatic Urdu paraphrase types identification. The proposed corpus constitutes a valuable linguistic resource for Urdu natural language processing and provides empirical evidence of the effectiveness of multilingual transformer models for fine-grained paraphrase types identification in low-resource languages. Although the current corpus comprises annotated sentence pairs collected from the COUNTER corpus, online Urdu news articles, and manually curated examples, future work will extend the corpus to include additional textual domains and larger datasets for broader coverage of Urdu paraphrase types identification phenomena.
Source: Scientific Reports
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $2,690
OpenAlex ID: W7226018522
