ParaCotta: Synthetic Multilingual Paraphrase Corpora from the Most Diverse Translation Sample Pair

ParaCotta: Synthetic Multilingual Paraphrase Corpora from the Most Diverse Translation Sample Pair Alham Fikri Aji∗• Tirana Noor Fatyanosa∗◦ Radityo Eko Prasojo∗⋆ Philip Arthur† Suci Fitriany∗ Salma Qonitah∗ Nadhifa Zulfa∗ Tomi Santoso∗ Mahendra Data‡◦ ∗ Kata.ai Research Team, ◦ Kumamoto University, • University of Edinburgh, ⋆ Universitas Indonesia, † Oracle Digital Assistant, ‡ Brawijaya University {aji,tirana.fatyanosa,ridho,suci}@kata.ai {salma.qonitah,dhifa.zulfa,tomi.santoso}@kata.ai {fatyanosa, mahendra.data}@dbms.cs.kumamoto-u.ac.jp philip.arthur@oracle.com, mahendra.data@ub.ac.id Abstract We release our synthetic parallel paraphrase corpus across 17 languages: Arabic, Catalan, Czech, German, English, Spanish, Estonian, French, Hindi, Indonesian, Italian, Dutch, Ro- manian, Russian, Swedish, Vietnamese, and Chinese. Our method relies only on mono- lingual data and a neural machine translation system to generate paraphrases, hence simple to apply. We generate multiple translation sam- ples using beam search and choose the most lexically diverse pair according to their sen- tence BLEU. We compare our generated cor- pus with the ParaBank2. According to our evaluation, our synthetic paraphrase pairs are semantically similar and lexically diverse. 1 Introduction Paraphrases are semantically similar sentences or phrases using different expressions (Bhagat and Hovy, 2013). A paraphrase generation system can be developed by training a model, given a dataset of paraphrase parallel texts (Egonmwan and Chali, 2019). Paraphrase parallel corpus is accessible in English (Dolan and Brockett, 2005; Fader et al., 2013; Xu et al., 2015). However, such data for other languages are not as common. (Ganitkevitch and Callison-Burch, 2014) proposed a multilingual para- phrase pairs dataset; however, their corpus is only on phrase-level. By utilizing a machine translation system, Wieting and Gimpel (2017) proposed synthetic paraphrase corpus by back-translating bilingual text. However, this approach does not consider lexical diversity for the generated paraphrases. Hu et al. (2019a) and Hu et al. (2019b) further improve the method by apply- ing a constraint to generate more diverse paraphrases, then choosing diverse pairs as the synthetic dataset. These methods require bilingual corpus, which might not be easily accessible for certain languages or do- mains. Our work takes inspiration from selecting the most diverse pair; however, we remove the bilingual text requirement. We propose a simple way to generate paraphrases by selecting the most diverse pair (in terms of BLEU) from the translation sample. Our approach generates paraphrases from a monolingual text; therefore, not bound to the availability of parallel corpus. We also show that this technique can produce diverse para- phrases, measured in the BLEU score. In addition, we release our generated paraphrase dataset in 17 languages. 2 Generating Paraphrase via Diverse Pairs We propose a way to construct synthetic paraphrase corpus by utilizing a machine translation system. Our approach involves translating texts from En- glish to the desired language, therefore not limited to the availability of bilingual corpora for back- translation (Hu et al., 2019a). Specifically, given Never gonna let you down. Monolingual English corpus: Multiple translations: Jangan pernah mengecewakanmu. Tidak akan pernah mengecewakanmu. Tak akan pernah mengecewakanmu. Takkan pernah mengecewakanmu. Tidak akan pernah mengecewakan Anda. Tidak akan mengecewakanmu. Jangan pernah mengecewakanmu. | Tidak akan pernah mengecewakan Anda. Select 2 most different text as …

Content truncated to 500 words