A small research of interaction between span-based Model from Ouchi et al. (2018) and types of contextualized embeddings (focused on ELMo, BERT and stacked Embeddings between Flair and BERT) with SENNA Embeddings as an example of baseline. All datas used come from CoNLL 2012.
├── scripts
│ ├── create_embedding - create the contextualized embeddings used in experiments
│ ├── plot - create 3 types of plots for results in experiment (scatter plot, imshow and line plot)
│ ├── prepare - prepare the raw datas for embeddings training and gold standards for evaluation
└── results - Results for each types of embeddings including visualizations, evaluations for results at different epochs for test set, training report (output from this model)
│ ├── bert
│ ├── elmo
│ ├── senna
│ └── stacked_bert-flair
Note: all train reports for BERT and Stacked have the same name types of ELMo as Ouchi only offers 2 types of trained embeddings: non-contextualised and ELMo (on behalf of contextualised) embeddings.
- Report in German written for the results created in October 2019.