Re-TACRED: Addressing Shortcomings of the TACRED Dataset
George Stoica, Emmanouil Antonios Platanios, and Barnabás Póczos
In Proceedings of the Thirty-fifth AAAI Conference on Artificial Intelligence 2021

Primary Contact: George Stoica. As of Jan 2021, I am no longer at CMU, and the cs.cmu.edu email may no longer work. Please contact me instead at: gstoica27@gmail.com.

Changelog

• 1.0 - Initial dataset release: Data consisted of 105,206 total instances spread across 40 relations.
• 1.1 - Updated dataset release: After extensive discussion, we have elected to prune Re-TACRED by ~ 14K instances. The new dataset has 91,467 instances, spread across 40 relations. Pruned data consisted of a mixture of messily segmented entities (and corresponding types), or sentences whose relations were ambigious. While this version is smaller, it is cleaner, and better defined.

This repository contains all relevant resources for using Re-TACRED, a new relation extraction dataset.

For details on this work please check out our:

• arXiv: Paper
• AAAI 2021: Paper & Poster
• NeurIPS 2020 KR2ML Workshop: Paper & Poster

Below we describe the contents of the four repository directories by name.

Re-TACRED

This directory contains version 1.1 of our revised TACRED dataset patches for each split. Due to licensing restrictions, we cannot provide the complete dataset. However, following Alt, Gabryszak, and Hennig (2020), our patch consists of json files mapping TACRED instances by their id to our revised labels.

The original TACRED dataset is available for download from the LDC here. It is free for members, or $25 for non-members.

Applying the patch is simple and only requires replacing each TACRED instance (where applicable) with our revised relation. For convenience, we provide a script for this named apply_patch.py in the Re-TACRED directory. In the script, you only need to replace

tacred_dir = None
save_dir = None

With the path to your TACRED dataset save directory, and the directory where you wish to save the patched data to respectively.

PA-LSTM, C-GCN & SpanBERT

We base our experiments off of the open-source model repositories of:

• PA-LSTM: Zhang et. al. (2017)
• C-GCN: Zhang et. al. (2018)
• SpanBERT: Joshi et. al. (2019)

However, it is not possible to simply pass Re-TACRED to each model repository because each is hardcoded for TACRED. Thus, we must modify certain files to make each model Re-TACRED compatible. To make it as easy as possible, we provide all our altered files in each named model directory (e.g., the provided PA-LSTM directory). All that needs to be done is to replace the corresponding file in our provided directory with the corresponding file in the original model repository. For instance, you may replace SpanBERT's "run_tacred.py" file with our "run_tacred.py" file. Running experiments is equivalent to how it is performed in the original model repositories.

Note that our files also contain certain "quality of life" changes that make running each model more convenient for us. Examples include adding and tracking the test split while training (as opposed to only the dev set).