Extend the DataFrame visualization from #203 to support making edits to the DataFrame itself to support data cleaning and active learning applications. Major types of edits that would be useful:
Change the begin and end offsets of spans by dragging
Add new spans (adding new rows to the DataFrame)
Remove spans (remove rows from the DataFrame)
Add labels to span (add a new column to the DataFrame with metadata like "is correct", or "entity type", and populate that column from the UI)
With this editing support, it should be possible to do a data cleaning application end-to-end, entirely in the notebook. Early cells in the notebook load the data, maybe train a model on it, and generate a DataFrame with information about what was found in each document. Then the user edits the DataFrames in place. Then later cells in the notebook consume the results from the editing session and do things like retraining the model or writing out a new, corrected data set.
This kind of interactivity and interplay between the JavaScript and the backing Python objects will require a JupyterLab widget.
frreiss
commented
2 years ago
Extend the DataFrame visualization from #203 to support making edits to the DataFrame itself to support data cleaning and active learning applications. Major types of edits that would be useful:
With this editing support, it should be possible to do a data cleaning application end-to-end, entirely in the notebook. Early cells in the notebook load the data, maybe train a model on it, and generate a DataFrame with information about what was found in each document. Then the user edits the DataFrames in place. Then later cells in the notebook consume the results from the editing session and do things like retraining the model or writing out a new, corrected data set.
This kind of interactivity and interplay between the JavaScript and the backing Python objects will require a JupyterLab widget.