PoetryDiffusion: Towards Joint Semantic and Metrical Manipulation in Poetry Generation
Authors: Zhiyuan Hu, Chumin Liu, Yue Feng, Anh Tuan Luu, Bryan Hooi
AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results on two datasets demonstrate that our model outperforms existing models in terms of automatic evaluation of semantic, metrical, and overall performance as well as human evaluation. |
| Researcher Affiliation | Academia | Zhiyuan Hu1*, Chumin Liu2, Yue Feng3, Anh Tuan Luu2, Bryan Hooi1 1National University of Singapore (NUS), Singapore 2 Nanyang Technological University (NTU), Singapore 3 University College London (UCL), UK |
| Pseudocode | No | The paper does not contain structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | Codes are released to https://github.com/Chorling Lau/Poetry Diffusion/. |
| Open Datasets | Yes | We train our model on two datasets, Sonnet and Song Ci. Sonnet consists of 3,355 sonnets collected by (Lau et al. 2018). Song Ci comprises 82,724 Song Ci s, curated by (Zhang et al. 2019). |
| Dataset Splits | Yes | To ensure a fair comparison, datasets in two languages are partitioned into train/valid/test in the same way as used in previous work. |
| Hardware Specification | Yes | It takes approximately 4 hours to train Poetry Diffusion and Metrical Controller on an NVIDIA A100 GPU monopolized by one job. |
| Software Dependencies | No | The paper mentions various models and techniques (e.g., BERT, GPT3, DDIM, LoRA) but does not provide specific version numbers for software dependencies like programming languages or libraries. |
| Experiment Setup | Yes | The number of decoding or encoding steps T is set to be 2000 steps. In addition, we rescale the diffusion steps into 200 to accumulate the poetry generation process based on DDIM (Song, Meng, and Ermon 2020). The dimension of word embedding is chosen to be 16. The method of organizing batches differs between the two datasets. For Sonnet, pad each piece of poetry to the same length and then concatenate the number of sequences corresponding to batch size. While for Song Ci, firstly concatenate all sequences of text and then cut into blocks with appropriate shapes. The number of training iterations is set to 150K. |