Learning to Write Stories with Thematic Consistency and Wording Novelty
Authors: Juntao Li, Lidong Bing, Lisong Qiu, Dongmin Chen, Dongyan Zhao, Rui Yan1715-1722
AAAI 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results on ROCStories and Writing Prompts indicate that our proposed model can generate stories with consistency and wording novelty, and outperforms existing models under both automatic metrics and human evaluations. |
| Researcher Affiliation | Collaboration | 1Center for Data Science, Academy for Advanced Interdisciplinary Studies, Peking University, Beijing, China 2Institute of Computer Science and Technology, Peking University, Beijing, China 3R&D Center Singapore, Machine Intelligence Technology, Alibaba DAMO Academy |
| Pseudocode | No | The paper does not contain structured pseudocode or algorithm blocks (clearly labeled algorithm sections or code-like formatted procedures). |
| Open Source Code | No | The paper does not provide concrete access to source code for the methodology described, such as a specific repository link, an explicit code release statement, or code in supplementary materials. |
| Open Datasets | Yes | To train our story generation model, we conduct experiments on two corpora: the ROCStories 2, and the Writing Prompts dataset. Specifically, the ROCStories corpus is created for the shared-task of Story Cloze Test (Cai, Tu, and Gimpel 2017; Schwartz et al. 2017), which is man-made with the following two merits: 1) It captures a rich set of commonsense relations of daily life; 2) It is a high-quality collection of life stories which can be used to learn story understanding and generation. As presented in Figure 3, each story of ROCStories comprises of exactly five sentences. The Writing Prompts dataset is collected from Reddit s Writing Prompts forum 3 for hierarchical story generation (Fan, Lewis, and Dauphin 2018). |
| Dataset Splits | Yes | To preprocess ROCStories, we first applied NLTK for tokenization, and then we randomly split the data into 78,527/9,816/9,816 stories for training/validation/test. For Writing Prompts, we followed (Fan, Lewis, and Dauphin 2018) for preprocessing: The dataset is randomly split into 272,600/15,138/15,620 stories as training/validation/test sets |
| Hardware Specification | No | The paper does not provide specific hardware details (exact GPU/CPU models, processor types with speeds, memory amounts, or detailed computer specifications) used for running its experiments. |
| Software Dependencies | No | The paper mentions using NLTK for tokenization but does not provide specific version numbers for NLTK or any other software libraries or solvers needed to replicate the experiment. |
| Experiment Setup | Yes | Both encoder and decoder are formed by one layer. The hidden states dimension of encoder and decoder are both set to 500. The word embedding size is 300 and shared across everywhere. The vocabulary size is comprised of most frequent 30,000 words. The Cache size is set to {20, 30, 50, 70} for the ROCStories and {50, 100, 200, 300} for the Writing Prompts corpus. The size of the latent variable z is 300. The prior network consists of 1 hidden layer with dimension of 400 and tanh activation function. All initial weights are uniformly sampled from [ 0.08, 0.08]. The batch size is set to 80. We use Adam optimizer (Kingma and Ba 2014) with learning rate of 0.001 and gradient clipping of 5 to train our models in an end-to-end fashion. |