Tree Labeller

The command line tool that helps label all leaves of a tree based only on a small sample of manually labelled leaves.

Sample scenarios include:

• Assigning shop departments to products organized in a taxonomy of categories
• Mapping taxonomy of book categories in one library to flat vocabulary of book categories in another library

This can be helpful when migrating data from one e-commerce system to another or preparing labelled data for machine learning project.

Here's example of the first task:

Labelling with the tool is a semi-automatic iterative process. You start by labeling few samples and the rule-based prediction algorithm tries to learn and tag the rest of the data set for you. You then correct predicted labels for a sample of most ambiguous items and the algorithm repeats prediction based on labels you provided. The algorithm suggests the most diverse sample of items to label, i.e. coming from different categories, so you don't waste time with samples that have high chance of having same label.

Install

Prerequisites: Python 3.8, 3.9.

To install the library:

pip install git+https://github.com/dzieciou/tree-labeller.git
export PATH=$PATH:$HOME/.local/bin

Usage

Describe your taxonomy in form of YAML file, e.g.: tree.yaml

name: categories
id: 1
children:
- name: Alcoholic Drinks
  id: 11
  children:
  - name: Whiskies
    id: 111
    children:
    - name: Jack Daniel's
      id: 1111
    - name: Johnnie Walker's
      id: 1112
  - name: Wines
    id: 112
    children:
    - name: Cabernet Sauvignon
      id: 1121
  - name: Beers
    id: 113
    children:
    - name: Guinness
      id: 1131

Create labelling task:

create_task \
    --dir ./my_labels \
    --tree ./tree.yaml \
    --allowed-labels Label1,Label2,Label3

Generate a sample:

label --dir ./my_labels --sample 10

Annotate a file with samples.

Run predictions and generate another sample of ambiguous and non labeled items. Items in the sample are sorted starting from the most ambiguous ones, i.e., having many possible label candidates.

label ./my_labels --sample 10

Repeat the process until you are satisfied.

After each iteration you will get statistics to help you decide when to stop labelling:

Iteration    Manual    Univocal    Ambiguous    Missing    Total    Allowed Labels

----------- -------- ---------- ----------- --------- ------- ---------------- 1 0 0% 0% 100% 14456 0%

2 10 71% 29% 0% 14456 37%

In the ideal situation we want to have 100% of univocal predictions, 0% of ambiguous and missing predictions and 100% of allowed labels (departments) coverage while providing as few manual labels as possible.

If you decide to continue, you can do one or more of the following actions:

• Correct ambiguous predicted labels in a sample.
• Correct your previous manual labels.
• Label with ? to skip the product from the prediction (it won't be sampled next time).
• Label with ! to tell the algorithm that the product ,and perhaps its category, are not present in the target shop (the algorithm will try to learn other similar products that might be not present in a shop)
• If one of departments have no products labeled so far, you can search for matching products manually and add them to the sample with correct label. For search you can use last TSV file with univocal predicted labels.
• You can also occasionally review univocal predicted labels and correct them by adding to the sample.

Task artifacts

In the labelling task directory the following artifacts are generated:

Filename	Description
config.yaml	Labeling task configuration.
tree.yaml	Taxonomy to label.
[n]-to-verify.tsv	Taxonomy leaves selected after n-th iteration for labelling/verification.
[n]-good.tsv	Taxonomy leaves with non-ambiguous labels predicted after n-th iteration.
[n]-mapping.tsv	Maps taxonomy categories (inner nodes) to labels after n-th iteration.
[n]-stats.json	Labeling statistics after n-th iteration.
all-stats.jsonl	Sequence of all iterations statistics accumulated so far.

Documentation

• Demo
• Sampling taxonomy leaves for manual labelling
• Distributing labelling budget
• Predicting labels

Development

Install poetry:

curl -sSL https://install.python-poetry.org | python3 -

Install dependencies:

poetry install

Activate virtual environment:

poetry shell

Install locally to test scripts:

deactivate
poetry build
pip install dist/tree_labeller-0.1.0-py3-none-any.whl

Acknowledgements

I would like to thank to:

• members of Computer Science Stack Exchange for help in developing algorithms
• my girlfriend Renata for help in mapping our local grocery store
• my colleagues from Samsung R&D Poland for feedback.