UP

• Collection of files for working with NER models
• Author: Juraj Dedič

• Scripts needed for conversion to SpaCy DocBin
• Also JSON converter for other training pipelines

• training of BERT model using HuggingFace

• evaluation for BERT / RoBERTa

• training RoBERTa using PyTorch
• was not getting better results than BERT
• so not used anymore

• Evaluation of SpaCy NER model

• Training of SpaCy NER and SpanCat models
• The training config was changed many times but the best was described in the thesis

• The data formats for training NER and SpanCat models are different
• this file converts the data to train SpanCat

• Evaluation of SpanCat model

• JuPyter notebooks were used mostly to work with the models
• I personally don't have the hardware to train the models
• Kaggle was used and it supports the Jupyter notebooks
• While working with the models there were tutorials used to help me create them (linked in the thesis)
• The first thing implemented was the data conversion script
• It converts the XML files with human transcripts to SpaCy DocBin
• Next the spacy NER model was trained
• There were multiple iterations of the models and the datasets changed
• There were also models BERT & RoBERTa trained using Pytorch and HuggingFace,
• These models were not used later, because it was found out that the SpaCy NER pipeline outperformed them
• After that the SpaCy spancat model was trained
• And the detected classes changed to include also the commands and values