.png){kind=icon}
.png){kind=icon}
Return
RAR: Recombination and augmented replacement method for insertion-based lexically constrained text generation
.png){kind=icon}
PRE
AI
DOI：10.1016/j.neucom.2024.127985.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Lexically constrained text generation aims to generate text by indicated keywords. Previous work has employed non-autoregressive insertion methods which iteratively insert tokens between keywords to generate complete sentences. However, the semantic constraints imposed by discrete tokens limit the flexibility of generated text, leading to a lack of diversity and regularity in the generated text. To address this issue, we introduce the RAR model which builds upon the iterative insertion -based generation method. This model uses recombination and augmented replacement method to change the structure and regularity of the text to enhance text quality, the RAR model offers flexibility by adjusting the positional relationships between tokens by recombination method. Meanwhile, we use two kinds of custom token -level classifiers to divide the Replacement method into two parts: hard replacement and soft replacement. The Hard Replacement'' involves substituting words to improve text structure, while the Soft Replacement'' focuses on modifying text regularity through word form adjustments. We adopt different replacement methods according to the different stages of text generation, bringing more flexibility to the generation process. Experimental results on multiple datasets demonstrate that RAR achieves significant improvements in text fluency and diversity.
Keywords:
Text generation
Lexically constrained
Replacement
Recombination
Journal
IF:
6.5
Papers:
2.5W
Citations:
6.5W