This project reviews the standard methods in text tagging and experiments extending the approach proposed in Universal Language Model Fine-tuning for Text Classification (ULMFiT) while integrating the modifications in a local copy of the FastAI library.
The files/folder contained in this repo are:
-
fastai/ directory: Contains the modified verion1.0.31of the fastai library to inlcude text tagging. -
ULMFiT_approach: A notebook with an execution of the Labeler (on working results) and some of the functions integrated in the library. -
Data_preprocessing_visualization_new.ipynb: A notebook with Data preprocessing, visualization for the presentation -
final_project_checkin_template.ipynb: first machine learning model fitting -
baseline_optimization.ipynb: A notebook with grid search and pipeline to tune the machine learning algorithms
The extension of the approach proposed in ULMiT to this task is still an ongoing project. While a working version has been constructed the models results still need to be improved.
While developing the application of ULMFit to text tagging we realized a major issue of using pre-defined models for that task. This is, the tokenization of the up-stream task, which generally is used for several down-stream tasks, needs to match the one that was provided in the down-stream taks for the text to match the labels.
Miguel Romero, Louise Lai, Jenny Kong.