Metal, a Multi-task Learning Framework for Sequence Modeling

Metal is a multi-task learning framework inspired by Collobert and Weston’s famous paper, but adapted for sequence modeling. It comes with its own simple domain-specific language (DSL) that allows you to ramp up sequence models for various NLP tasks (e.g., part-of-speech tagging, named entity recognition, semantic role labeling), without writing any actual code. Metal supports the expected deep learning components for sequence modeling, e.g., biLSTMs, CRF layers, character-level encoders, etc.

Metal is implemented on top of DyNet, using our own Scala packaging in the fatdynet project.


Table of contents