Paraphrase Types for Generation and Detection

TL;DR

The paper tackles the limitation of binary paraphrase tasks by proposing Paraphrase Type Generation and Paraphrase Type Detection to explicitly model linguistic perturbations. It formalizes the tasks around a taxonomy of paraphrase types $L$ and segment positions $s_i$, and evaluates using BLEU, ROUGE, and BERTScore, among other metrics. Through ETPC and auxiliary datasets (QQP, TURL, PAWS), it shows that models struggle with fine-grained paraphrase types, yet learning these types yields benefits for both generation and detection in traditional tasks, with scaling further enhancing performance. The work demonstrates cross-task transfer, provides correlations between type groups, and offers a demo and publicly available resources to advance this new paradigm toward more interpretable and robust paraphrase modeling.

Abstract

Current approaches in paraphrase generation and detection heavily rely on a single general similarity score, ignoring the intricate linguistic properties of language. This paper introduces two new tasks to address this shortcoming by considering paraphrase types - specific linguistic perturbations at particular text positions. We name these tasks Paraphrase Type Generation and Paraphrase Type Detection. Our results suggest that while current techniques perform well in a binary classification scenario, i.e., paraphrased or not, the inclusion of fine-grained paraphrase types poses a significant challenge. While most approaches are good at generating and detecting general semantic similar content, they fail to understand the intrinsic linguistic variables they manipulate. Models trained in generating and identifying paraphrase types also show improvements in tasks without them. In addition, scaling these models further improves their ability to understand paraphrase types. We believe paraphrase types can unlock a new paradigm for developing paraphrase models and solving tasks in the future.

Figures (9)

• Figure 1: Comparison of current paraphrase tasks (left) and our proposal towards paraphrase types (right).
• Figure 2: Paraphrase Type Generation (left) and Paraphrase Type Detection (right) with model $f$, reference segments $s1, s2, s3$ and candidate segments $\tilde{s}_1$, $\tilde{s}_2$, $\tilde{s}_3$.
• Figure 3: Task performance (accuracy) for different model sizes of LLaMA with and without learned paraphrase types against human performance as reported by the respective datasets or benchmarks.
• Figure 4: Accuracy for identifying paraphrase types of six high-level groups for different model sizes in the ETPC dataset using LLaMA.
• Figure 5: Rescaled Spearman correlations between paraphrase types in six higher-level families using word position deviation and lexical deviation. Correlations are normalized around the origin using mean $\mu$ and standard deviation $\sigma$ such that $\mu=0, \sigma=1$.