qfirst

Working repository for QA-SRL research. Much of the contents of this repo have ended up in qasrl-modeling or qasrl-roles.

Contents

• qfirst/: Python directory with ML model code.
• qfirst-scala/: Scala directory with other utilities:
  • clause/: Automatically convert QA-SRL to the simple clausal format.
  • clause-ext/: An extended clausal format including two preposition slots.
  • clause-ext-ann/: Annotate clause disambiguation in the extended format.
  • clause-ext-demo/: Show the construction of clause-level output from a question-first model trained with the extended clausal format.
  • frame/: Induce and evaluate verb frames.
  • frame-ann/: Manually annotate verb frames.
  • frame-browse/: View induced frames and construct gold data for question-paraphrasing eval.
  • metrics/: General utilities for computing metrics.
  • model-eval/: Evaluate QA-SRL models with varying output formats.
  • model-gen/: Auto-generate AllenNLP configs for all of the QA-SRL model components.
  • reprocess/: Fix the slot formatting errors in the QA-SRL Bank 2.0.

Setup

Run ./scripts/setup.sh to download extra data files.

Python setup

I need to document this. I can't remember if AllenNLP was used as a source or library dependency, and what version, or what version of PyTorch. The rest is in requirements.txt though.

Scala setup

You need the mill build tool version 0.5.1 or later.

Usage

Still working on fixing up and documenting usage for all of the modules.

Clause formation / discourse question demo

After doing the setup, you should have mill installed and you should have a file data/clausal-predictions.jsonl of model predictions. Then, in scripts/clause_demo.scala, modify the domain and httpPort values to the ones you want to host the demo at (localhost is fine for the domain). Then execute ./scripts/run_clause_demo.sh and after it finishes printing output the demo will be running at http://<domain>:<port>/task/evaluation/preview.