metadata
license: mit
HuffPost Relation-Enriched Dataset for Machine Learning
Description
This dataset is a semantically enriched version of the original HuffPost News Category dataset. It has been specifically re-engineered Machine-Learning tasks.
Unlike the original dataset, this version incorporates explicit semantic relationships and concept embeddings to help models generalize better with very few examples.
Key Enhancements
- Relation-Centric Formatting: Each entry is structured to support the
[ENTITY] [IS_A] [CATEGORY]relationship. - Semantic Expansion: Includes enriched context derived from semantic analysis (WordNet/Lesk integration) to provide better conceptual grounding.
- Concept Vector Alignment: Prepared for use with hybrid embedding architectures (e.g., combining MiniLM/BERT with specialized concept vectors).
Intended Use
This dataset is designed for researchers working on:
- Deep Learning Algorithms: Specifically for fast adaptation in NLP.
- Fusion Models: Testing how dynamic fusion models balance text vs. conceptual information.
- Relation Extraction: Benchmarking models on news-domain hierarchical relationships.
Technical Justification for GPU Grant
The enrichment process and the subsequent training on this dataset using ML involve complex gradient tracking and high-dimensional fusion (Dynamic Gated Fusion). This requires significant VRAM to handle the inner-loop adaptation across multiple tasks.
Sources
- Original Dataset: rmisra/news-category-dataset
- Modifications: Semantic enrichment and relation-centric pooling by [Ton-Pseudo]