EPARAPHRASING OF URDU TEXT USING NATURAL LANGUAGE PROCESSING

Abstract

Natural language processing is the walk through gate to interact with the computer through the natural languages which are spoken commonly. Paraphrasing of a text is basically the conversion of certain text in such a way that it’s semantic or meaning doesn’t change. We proposed an approach for paraphrasing of Urdu text which is a low constraint language with less data set and libraries. To deal with this process we divided our task into two sub-tasks which are i) re-ordering of the words in the sentence and ii) Changing the words with their appropriate synonym. Re-ordering of the words is done using the BART model which is a denoising sequence to sequence pre-trained model. We collected our own data set which contains the original and paraphrased Urdu sentences manually typed by the human. The BART model was trained on this data set. Bart is the bidirectional auto encoder which deals with the task of changing the order of the words along with the fill in novel spaces according to the grammar. The output is then passed to synonym replacement model which also have a separate data set which was collected by us. It contains the words with their synonyms and these words are replaced by their particular synonyms. So, we integrated both the models to get the desired result which are the paraphrasing of an Urdu text. The experiment shown that our model performed quite well, and the results were as desirable. We evaluated our model based on BLEU score. The BLEU score for the predicted text was ”0.54” when compared to the human paraphrased text.