hmtl

Why we included this project

This repository ships the reference implementation behind an AAAI 2019 paper on hierarchical multi-task learning, where one shared architecture handles several NLP tasks instead of a separate model per job. The trick is the layering: simpler semantic tasks like named entity recognition and entity mention detection sit at the bottom and feed into harder ones such as relation extraction and coreference resolution, with the authors reporting state-of-the-art results on three of those four. For NLP researchers and graduate students, the main value is as a compact, readable template rather than a ready-to-deploy product. Training, fine-tuning, and evaluation code is all present, alongside pre-trained weights and an online demo, and because it is built directly on PyTorch and AllenNLP, anyone comfortable with those libraries can adapt the architecture to their own task combinations.