Hierarchical Multilabel Document Classification
Hierarchical multi-label document categorization using a label taxonomy (tree/DAG) and neural + classical baselines.
Overview
This project focuses on hierarchical multi-label document classification (HMLC) — assigning multiple labels to each document, where labels are organized in a hierarchy (e.g., Category → Subcategory → Group → Subgroup). Unlike “flat” multi-label classification, HMLC explicitly models parent–child relations between categories.
- Task: multi-label document classification with a predefined taxonomy
- Output: one or more paths in the hierarchy (multiple labels per document)
How it works
A hierarchical label structure has been emplyed to classify documents into their appropriate categories with the goal to explore:
- how hierarchical constraints affect prediction quality,
- how multi-level labels can be represented (per-level indices vs. global label ids),
- and how different models behave when labels are not independent.
Data & label structure (example)
Documents are represented with text fields (e.g., title/abstract) and assigned (hierarchical) labels.
A typical record contains tokenized text plus labels at each level and a combined list of label ids:
-
title,abstract: tokenized text -
section / subsection / group / subgroup: indices at each hierarchy level -
labels: encoded into label ids (using per-level offsets)
Modeling
Approaches explored/used (depending on experiment scope):
- Hierarchy-aware neural model (e.g., attention + recurrent encoder with hierarchical output heads)
- Multi-label baselines (flat vs. hierarchical)
- Evaluation with hierarchy in mind (per-level accuracy/F1 and overall multi-label metrics)
Why hierarchy helps
A hierarchy provides:
- regularization (predictions are constrained by parent labels),
- better interpretability (you can explain predictions as paths),
- and a natural way to handle fine-grained label spaces.
