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 Syntax Attention-Guided Paraphrase (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. |
Year | DOI | Venue |
---|---|---|
2022 | 10.1016/j.neucom.2022.02.020 | Neurocomputing |
Keywords | DocType | Volume |
Syntax-controlled paraphrasing,Semantic consistency,Syntactic controllability,Coarse-grained Syntactic Control | Journal | 485 |
ISSN | Citations | PageRank |
0925-2312 | 0 | 0.34 |
References | Authors | |
0 | 7 |
Name | Order | Citations | PageRank |
---|---|---|---|
Erguang Yang | 1 | 0 | 0.68 |
Mingtong Liu | 2 | 0 | 0.34 |
Deyi Xiong | 3 | 845 | 67.74 |
Yujie Zhang | 4 | 251 | 52.63 |
Yao Meng | 5 | 0 | 1.69 |
Jin An Xu | 6 | 15 | 24.50 |
Yufeng Chen | 7 | 38 | 16.55 |