Abstract:As the neural-based Seq2Seq model pushes the state-of-the-art in text generation, recent work has turned to controlling attributes of the text such models generate, where syntax-controlled text generation can be applied for the paraphrase generation task, i.e., given an input sentence and a syntactic control to generate a paraphrase. The main challenge is how to generate sentences that follow the given syntax while maintaining semantics during the decoding process. Previous approaches using constituency parse trees as syntactic control can achieve better performance, but still suffer from the problems of inaccurate utilization of syntactic information and original semantics loss. To this end, we propose a S yntax A ttention- G uided P araphrase (SAGP) generation model that can utilize the previously generated text to accurately select a syntactic node from the given constituency parse tree to guide the generation of paraphrases. The automatic and manual evaluation results on the public datasets of syntactically controlled paraphrase generation task show that SAGP achieves state-of-the-art results in both syntactic controllability and semantic consistency. In order to improve the semantic consistency, we further propose a coarse-grained syntactic control definition method, which first removes the part-of-speech node and then extracts higher-level subtrees as control, so that meaningful paraphrases can be generated within a loose constraint. The experimental results on the same evaluation set show that the coarse-grained syntactic control can significantly improve semantic consistency.

An Adversarial Approach for Unsupervised Syntax-Guided Paraphrase Generation

ParaMac: A General Unsupervised Paraphrase Generation Framework Leveraging Semantic Constraints and Diversifying Mechanisms.

Syntactically-Informed Unsupervised Paraphrasing with Non-Parallel Data.

Adversarial Example Generation with Syntactically Controlled Paraphrase Networks

Unifying Discrete and Continuous Representations for Unsupervised Paraphrase Generation

Improving Generation Diversity Via Syntax-Controlled Paraphrasing

Unsupervised Paraphrasing by Simulated Annealing

SECT: A Successively Conditional Transformer for Controllable Paraphrase Generation

A Submodular Optimization-Based VAE-Transformer Framework for Paraphrase Generation.

Exploring Bilingual Parallel Corpora for Syntactically Controllable Paraphrase Generation

MCPG: A Flexible Multi-Level Controllable Framework for Unsupervised Paraphrase Generation.

Exploring Diverse Expressions for Paraphrase Generation

A Learning-Exploring Method to Generate Diverse Paraphrases with Multi-Objective Deep Reinforcement Learning

Unsupervised Paraphrasing under Syntax Knowledge

Unsupervised Disentanglement Learning Model for Exemplar-Guided Paraphrase Generation

Paraphrase Generation As Unsupervised Machine Translation.

Learning Structural Information for Syntax-Controlled Paraphrase Generation.

Augmenting Paraphrase Generation with Syntax Information Using Graph Convolutional Networks.

A Semantically Consistent and Syntactically Variational Encoder-Decoder Framework for Paraphrase Generation.

A Quality-based Syntactic Template Retriever for Syntactically-controlled Paraphrase Generation

Diversity-Enhanced Learning for Unsupervised Syntactically Controlled Paraphrase Generation