Jupyter multi-label text classification widget, ideal for creating few-shot learning annotations

annotation_widget.py

# Annotator Widget
# Copyright (C) 2022 Pablo Duboue - Licensed under MIT license
# define the following variables beforehand:
#   classes = list of strings, three character classes display better
#   titles = list of strings, one title per document to be annotated
#   texts = list of (list of strings), one list of lines (strings) per document to be annotated
#   annotations = [ set() for _ in range(len(texts)) ] # annotations
#   current = 0 # current document being displayed
#   fulltext = False # whether to show full text or top/bottom
from ipywidgets import widgets
def widget():
    annotator = widgets.VBox()
    
    def render():
        title = widgets.Label(titles[current], layout=widgets.Layout(width="600px"))
        pos = widgets.Text(description='pos',value=str(current), layout=widgets.Layout(width="120px"))
        gobtn = widgets.Button(description='go', layout=widgets.Layout(width="40px"))
        def go(b):
            global current
            current = int(pos.value)
            render()
        gobtn.on_click(go)
        nxtbtn = widgets.Button(description='next', layout=widgets.Layout(width="60px"))
        def nxt(b):
            global current
            current = (current + 1) % len(texts)
            render()
        nxtbtn.on_click(nxt)
        header = widgets.HBox(children = (title,pos,gobtn,nxtbtn))

        vchildren = list()
        showtoggle = widgets.Button(description="Top/Bottom" if fulltext else "See All Text")
        def toggle(b):
            global fulltext
            fulltext = not fulltext
            render()
        showtoggle.on_click(toggle)
        vchildren.append(widgets.HBox(children = (
            widgets.Label("Lines: {:,}".format(len(texts[current]))), showtoggle)
        ))
        def toggleann(evt):
            global annotations
            global current 
            if evt['name'] == 'value':
                if evt['new']:
                    annotations[current].add(evt['owner'].description)
                else:
                    annotations[current].remove(evt['owner'].description)
        wclasses = classes
        while wclasses:
            children = [ widgets.Checkbox(
                description=c, value=c in annotations[current],
                layout=widgets.Layout(width="140px")
            ) for c in wclasses[:5] ]
            vchildren.append(widgets.HBox(children = children))
            wclasses = wclasses[5:]
            for c in children:
                c.observe(toggleann)
                
        if fulltext:
            full = widgets.Textarea(value="\n".join(texts[current]), rows=50) 
            vchildren.append(full)
        else:
            top = widgets.Textarea(value="\n".join(texts[current][:100]), rows=10)
            bottom = widgets.Textarea(value="\n".join(texts[current][-100:]), rows=10)
            vchildren.append(widgets.HBox(
                    children = (widgets.Label("Top"), top, widgets.Label("Bottom"), bottom)
            ))

        viewer = widgets.VBox(children=vchildren)
        annotator.children = ( header, viewer )
    
    render()
    return annotator

widget()            

Looks like this:

As you go annotating, the annotations dictionary gets populated and can be used concurrent with the annotation process.