Automated Paraphrase Quality Assessment Using Language Models and Transfer Learning

Learning to paraphrase supports both writing ability and reading comprehension, particularly for less skilled learners. As such, educational tools that integrate automated evaluations of paraphrases can be used provide timely feedback enhance learner paraphrasing skills more efficiently effectively. Paraphrase identification is a popular NLP classification task involves establishing whether two sentences share similar meaning. quality assessment slightly complex task, in which pairs are evaluated in-depth across multiple dimensions. In this study, we focus on four dimensions: lexical, syntactical, semantic, overall quality. Our study introduces evaluates various machine learning models using handcrafted features combined with Extra Trees, Siamese neural networks BiLSTM RNNs, pretrained BERT-based models, together transfer from larger general corpus, estimate the Two datasets considered tasks involving quality: ULPC (User Language Corpus) containing 1998 smaller dataset 115 based children’s inputs. The MSRP (Microsoft Research 5801 paraphrases. On dataset, our BERT model improves upon previous baseline by at least 0.1 F1-score When fine-tuning children network improve their original scores 0.11 F1-score. results these experiments suggest generic successful, while same time obtaining comparable fewer epochs.