STEPS: Semantic Typing of Event Processes with a Sequence-to-Sequence Approach
Authors: Sveva Pepe, Edoardo Barba, Rexhina Blloshmi, Roberto Navigli11156-11164
AAAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this Section we detail the experimental setting in which we train and evaluate our approach for Multi-axis Event Process Typing. In addition, we also describe the dataset we used, the sequence-to-sequence model we proposed alongside its training hyperparameters, and finally, our main comparison systems along with several STEPS variants. |
| Researcher Affiliation | Academia | Sapienza NLP Group, Sapienza University of Rome pepe.1743997@studenti.uniroma1.it, {barba,blloshmi}@di.uniroma1.it, navigli@diag.uniroma1.it |
| Pseudocode | No | No pseudocode or algorithm blocks were found in the paper. |
| Open Source Code | Yes | We release the data and software at https://github.com/Sapienza NLP/steps. |
| Open Datasets | Yes | To support the Multi-axis Event Process Typing task, Chen et al. (2020) constructed an automatic dataset by scraping the how-to guides from wiki How.3 Each page in wiki How corresponds to a complete guide divided into multiple steps on how to do something (e.g., the page How to Cook Pasta discusses the steps required to prepare pasta). ...We release the data and software at https://github.com/Sapienza NLP/steps. |
| Dataset Splits | Yes | Finally, following Chen et al. (2020), we train all the models on a split containing 80% of the whole dataset, and use as validation and test sets two equally sized partitions which include the remainder of the data. |
| Hardware Specification | Yes | The experiments are carried out using an Nvidia Ge Force RTX 2080ti. |
| Software Dependencies | No | STEPS builds on top of BART (Lewis et al. 2020), a sequence-to-sequence architecture pretrained with denoising objectives...We train STEPS with Adagrad (John, Elad, and Yoram 2011)...We use the pretrained weights of both models made available by the Transformers library (Wolf et al. 2020). While key software components are mentioned, specific version numbers (e.g., for the Transformers library, Python, PyTorch/TensorFlow) are not provided. |
| Experiment Setup | Yes | We train STEPS with Adagrad (John, Elad, and Yoram 2011) for a maximum of 300,000 steps with a learning rate of 2 * 10^-5, batches of 800 tokens and a gradient accumulation of 10 steps. Finally, we evaluate STEPS performances every 2,000 updates, and we interrupt the training if no improvements are observed in the validation set for 3 consecutive evaluations. At prediction time we use beam decoding with a beam size of 5. |