Results directory contents

This is the directory containing the results. The code will check to make sure this directory does not exist before trying to run anything, and ExperimentRunner uses the existence of the directory to know which experiments still need to be run.

Inside the main results directory, there is a settings.tsv file that basically lists all of the arguments and parameters that were used to run the code, so hopefully if you lose track of which output was in which directory, this file will help you figure it out.

Beyond that, there is one directory per relation tested. Inside that relation directory are files giving what paths were found, what paths were kept, what the learned path weights were, what the resultant feature matrices were that were used for training and testing, and what the final predictions were during testing. The matrices can be quite large, so I do not output them by default. You can output them by setting "output matrices": true at the top level of an experiment specification (this location for that particular parameter is not ideal, and is on my list of things to change).

The most important file for you to know about is probably the scores.tsv file. This file has all of the scores for each (source, target) pair found. Each source is listed in its own line chunk, with all found targets listed in the order they were predicted, with a corresponding score. There’s a * in the last column if the item was a known test example, and *^ if the item was a known training example.

To compute metrics over these results and compare experiments, use ExperimentScorer.