Event Representations for Automated Story Generation with Deep Neural Nets
Authors: Lara Martin, Prithviraj Ammanabrolu, Xinyu Wang, William Hancock, Shruti Singh, Brent Harrison, Mark Riedl
AAAI 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We give empirical results comparing different event representations and their effects on event successor generation and the translation of events to natural language. The data was split into 80% training, 10% validation, and 10% test data. All reported results were evaluated using the the held-out test data. |
| Researcher Affiliation | Academia | Lara J. Martin, Prithviraj Ammanabrolu, Xinyu Wang, William Hancock, Shruti Singh, Brent Harrison, Mark O. Riedl School of Interactive Computing Georgia Institute of Technology Atlanta, GA {ljmartin; raj.ammanabrolu; lillywang1126; whancock; shruti.singh; brent.harrison; riedl} @gatech.edu |
| Pseudocode | No | The paper does not contain structured pseudocode or algorithm blocks. Methods are described in prose. |
| Open Source Code | No | The paper does not provide a statement or link indicating that the source code for the described methodology is open-source or publicly available. |
| Open Datasets | Yes | For this work, we use a corpus of movie plot summaries extracted from Wikipedia (Bamman, O Connor, and Smith 2014) |
| Dataset Splits | Yes | The data was split into 80% training, 10% validation, and 10% test data. |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., GPU/CPU models, memory, or cloud instance types) used for running the experiments. It only mentions using TensorFlow and PyTorch as software frameworks. |
| Software Dependencies | Yes | Verbs were replaced by Verb Net (Schuler 2005) version 3.2.41 frames (e.g. arrived / arriving become escape-51.1 ). Events are created by first extracting dependencies with Stanford s Core NLP (Manning et al. 2014). We ran topic modeling on the entire corpus using Python s Latent Dirichlet Analysis2 set for discovering 100 different categories. The setup for this set of experiments is almost identical to that of the event2event experiments, with the main difference being that we used Py Torch3 which more easily lent itself to implementing beam search. |
| Experiment Setup | Yes | Each network was trained with the same parameters (0.5 learning rate, 0.99 learning rate decay, 5.0 maximum gradient, 64 batch size, 1024 model layer size, and 4 layers), varying only the input/output, the bucket size, the number of epochs and the vocabulary. The neural nets were trained until the decrease in overall loss was less than 5% per epoch. This took between 40 to 60 epochs for all experiments. The LSTM RNN networks in these experiments use beam search instead of greedy search to aid in finding a more optimal solution while decoding. The beam search decoder works by maintaining a number of partial hypotheses at each step (known as the beam width or B, where B=5). |