An Augmented Encoder to Generate and Evaluate Paraphrases in Punjabi Language

Abstract

Paraphrase generation is an important task in Natural Language Processing (NLP) and is successfully applied in various applications such as question-answering, information retrieval & extraction, text summarization and augmentation of machine translation training data. A lot of research has been carried out on paraphrase generation but in the language of English only. However, no approach is available for paraphrase generation in Punjabi Language. Hence, this paper aims to plug in the gap by developing a paraphrase generation and evaluation model for the language of Punjabi. The proposed approach is divided into two phases: paraphrase generation and evaluation. To generate paraphrases, the current state-of-the-art transformer with improved encoder is being used as transformers can learn long-term dependencies. For evaluation, the sentence embeddings are used to check whether the generated paraphrase is similar to the given sentence or not. The sentence embeddings have been created using two approaches: Seq2Seq with attention and transformers. The proposed model is compared with the currently available state-of-the-art models on Quora Question pair dataset. However, for Punjabi, the proposed approach is evaluated on three datasets: news headlines, the sentential dataset from news articles and the third dataset is the translation of Quora Question pair into Punjabi. The automatic evaluation metrics BLEU, METEOR and ROUGE are used for depth evaluation along with human judgments. The proposed approach is straightforward and successfully applies for augmenting machine translation training data and sentence compression. The proposed approach establishes a new baseline for paraphrase generation in Indian regional languages in the future.