LangHuAn

Language Human Annotations, a frontend for tagging AI project labels, drived by pandas dataframe data.

From Chinese word 琅嬛[langhuan] (Legendary realm where god curates books)

Here's a 5 minutes youtube video explaining how langhuan works

Installation

pip install langhuan

Minimun configuration walk through

langhuan start a flask application from pandas dataframe 🐼 !

Simplest configuration for NER task 🚀

from langhuan import NERTask

app = NERTask.from_df(
    df, text_col="description",
    options=["institution", "company", "name"])
app.run("0.0.0.0", port=5000)

Simplest configuration for Classify task 🚀

from langhuan import ClassifyTask

app = ClassifyTask.from_df(
    df, text_col="comment",
    options=["positive", "negative", "unbiased", "not sure"])
app.run("0.0.0.0", port=5000)

Frontend

You can visit following pages for this app.

Tagging

http://[ip]:[port]/ is for our hard working taggers to visit.

Admin

http://[ip]:[port]/admin is a page where you can 👮🏽‍♂️:

• See the progress of each user.
• Force save the progress, (or it will only save according to save_frequency, default 42 entries)
• Download the tagged entries

Advanced settings

Validation

You can set minimun verification number: cross_verify_num, aka, how each entry will be validated, default is 1

If you set cross_verify_num to 2, and you have 5 taggers, each entry will be seen by 2 taggers

app = ClassifyTask.from_df(
    df, text_col="comment",
    options=["positive", "negative", "unbiased", "not sure"],
    cross_verify_num=2,)

Preset the tagging

You can set a column in dataframe, eg. called guessed_tags, to preset the tagging result.

Each cell can contain the format of tagging result, eg.

{"tags":[
    {"text": "Genomicare Bio Tech", "offset":32, "label":"company"},
    {"text": "East China University of Politic Science & Law", "offset":96, "label":"company"},
    ]}

Then you can run the app with preset tag column

app = NERTask.from_df(
    df, text_col="description",
    options=["institution", "company", "name"],
    preset_tag_col="guessed_tags")
app.run("0.0.0.0", port=5000)

Order strategy

The order of which text got tagged first is according to order_strategy.

Default is set to "forward_match", you can try pincer or trident

Assume the order_by_column is set to the prediction of last batch of deep learning model:

• trident means the taggers tag the most confident positive, most confident negative, most unsure ones first.

Load History

If your service stopped, you can recover the progress from cache.

Previous cache will be at $HOME/.cache/langhuan/{task_name}

You can change the save_frequency to suit your task, default is 42 entries.

app = NERTask.from_df(
    df, text_col="description",
    options=["institution", "company", "name"],
    save_frequency=128,
    load_history=True,
    task_name="task_NER_210123_110327"
    )

Admin Control

This application assumes internal use within organization, hence the mininum security. If you set admin_control, all the admin related page will require adminkey, the key will appear in the console prompt

app = NERTask.from_df(
    df, text_col="description",
    options=["institution", "company", "name"],
    admin_control=True,
    )

From downloaded data => pytorch dataset

For downloaded NER data tags, you can create a dataloader with the json file automatically:

• pytorch + huggingface tokenizer
• tensorflow + huggingface tokenizer, development pending

Gunicorn support

This is a light weight solution. When move things to gunicorn, multithreads is acceptable, but multiworkers will cause chaos.

gunicorn --workers=1 --threads=5 app:app

Compatibility 💍

Well, this library hasn't been tested vigorously against many browsers with many versions, so far

• compatible with chrome, firefox, safari if version not too old.