Error Types in Transformer-Based Paraphrasing Models: A Taxonomy, Paraphrase Annotation Model and Dataset

Developing task-oriented bots requires diverse sets of annotated user utterances to learn mappings between natural language utterances and user intents. Automated paraphrase generation offers a cost-effective and scalable approach for generating varied training samples by creating different versions of the same utterance. However, existing sequence-to-sequence models used in automated paraphrasing often suffer from errors, such as repetition and grammar. Identifying these errors, particularly in transformer architectures, has become a challenge. In this paper, we propose a taxonomy of errors encountered in transformer-based paraphrase generation models based on a comprehensive error analysis of transformer-generated paraphrases. Leveraging this taxonomy, we introduced the Transformer-based Paraphrasing Model Errors dataset, consisting of 5880 annotated paraphrases labeled with error types and explanations. Additionally, we developed a novel multilabel paraphrase annotation model by fine-tuning a BERT model for error annotation task. Evaluation against human annotations demonstrates significant agreement, with the model showing robust performance in predicting error labels, even for unseen paraphrases.