TEDIN

TEDIN (Tree Edit DIstance Network) is a neural model for learning dynamic tree edit (TED) operation costs. It is implemented in a way to solve Natural Language Inference problems via TED.

This code includes a previous version of Infernal, but in order to use that system, it is advisable to refer to its dedicated repository.

Configuration

The file config.py contains variables with paths to some specific data files (e.g., wordnet, DELAF dictionary) and endpoints (stanford corenlp).

The code is somewhat hard coded to work with Portuguese data.

Usage

Before running any of the scripts, set the environment variable PYTHONPATH to include the current directory:

export PYTHONPATH=.

In order to use TEDIN, perform the following steps:

1. Run the preprocess.py scripts with XML files containing the data. It will need a running installation of the CoreNLP parser.
2. Run train-ranker.py with the trained data. This trains the model with distant supervision to learn operation costs.
3. Run train-classifier.py to train an actual NLI classifier from the model trained in the previous step.