Towards diverse and contextually anchored paraphrase modeling: A dataset and baselines for Finnish

Abstract In this paper, we study natural language paraphrasing from both corpus creation and modeling points of view. We focus in particular on the methodology that allows extraction challenging examples paraphrase pairs their textual context, leading to a dataset potentially more suitable for evaluating models’ ability represent meaning, especially document when compared with those gathered using various sentence-level heuristics. To end, introduce Turku Paraphrase Corpus, first large-scale, fully manually annotated paraphrases Finnish. The contains 104,645 labeled pairs, which 98% are verified be true paraphrases, either universally or within present context. order control diversity avoid certain biases easily introduced automatic candidate extraction, collected different paraphrase-rich text sources. This us create including longer lexically diverse than can expected through addition quality, also keep original context each pair, making it possible our knowledge, is provides pairs. several models trained evaluated new data. Our initial classification experiments indicate nature classifying detailed labeling scheme used annotation, accuracy substantially lacking behind human performance. However, large scale retrieval task almost 400M sentences, results highly encouraging, 29–53% being ranked top 10 depending type. Corpus available at github.com/TurkuNLP/Turku-paraphrase-corpus as well popular HuggingFace datasets under CC-BY-SA license.