ExplaGraphs

Dataset and PyTorch code for our EMNLP 2021 paper:

ExplaGraphs: An Explanation Graph Generation Task for Structured Commonsense Reasoning

Swarnadeep Saha, Prateek Yadav, Lisa Bauer, and Mohit Bansal

Website and Leaderboard

ExplaGraphs is hosted here. You can find the leaderboard, a brief discussion of our dataset, evaluation metrics and some notes about how to submit predictions on the test set.

Installation

This repository is tested on Python 3.8.3.
You should install ExplaGraphs on a virtual environment. All dependencies can be installed as follows:

pip install -r requirements.txt

Dataset

ExplaGraphs dataset can be found inside the data folder.

It contains the training data in train.tsv, the validation samples in dev.tsv and the test samples (without labels) in test.tsv.

Each training sample contains four tab-separated entries -- belief, argument, stance label and the explanation graph.

The graph is organized as a bracketed string (edge_1)(edge_2)...(edge_n), where each edge is of the form concept_1; relation; concept_2.

Evaluation Metrics

ExplaGraphs is a joint task that requires predicting both the stance label and the corresonding explanation graph. Independent of how you choose to represent the graphs in your models, you must represent the graphs as bracketed strings (as in our training data) in order to use our evaluation scripts.

We propose multiple evaluation metrics as detailed in Section 6 of our paper. Below we provide the steps to use our evaluation scripts.

Step 1

First, we evaluate the graphs against all the non-model based metrics. This includes computing the stance accuracy (SA), Structural Correctness Accuracy for Graphs (StCA), G-BertScore (G-BS) and Graph Edit Distance (GED). Run the following script to get these.

bash eval_scripts/eval_first

This takes as input the gold file, predictions file, and the relations file and outputs an intermediate file annotations.tsv. In this intermediate file, each sample is annotated with one of the three labels -- stance_incorrect, struct_incorrect and struct_correct. The first label denotes the samples where the predicted stance is incorrect, the second denotes the ones where the stance is correct but the graph is structurally incorrect and the third denotes the ones where the stance is correct and the graph is also structurally correct.

Structural Correctness Evaluation requires satisfying all the constraints we define for the task, which include the graph be connected DAG with at least three edges and having at least two exactly matching concepts from the belief and two from the argument. You SHOULD NOT look to boost this accuracy up by some arbitrary post-hoc correction of structurally incorrect graphs (like adding a random edge to make a disconnected graph connected).

Note that our evaluation framework is a pipeline, so the G-BS and GED metrics are computed only on the fraction of samples with annotation struct_correct.

Step 2

Given this intermediate annotation file, we'll now compute the Semantic Correctness Accuracy for Graphs (SeCA). Once again, this will only evaluate the fraction of samples where the stance is correct and the graphs are structurally correct. It is a model-based metric and we release our pre-trained model here. Once you download the model, run the following script to get SeCA.

bash eval_scripts/eval_seca

Step 3

In the final step, we compute the Edge Importance Accuracy (EA). This is again a model-based metric and you can download the pre-trained model here. Once you download the model, run the following script

bash eval_scripts/eval_ea

This measures the importance of an edge by removing it from the predicted graph and checking for the difference in stance confidence (with and without it) according to the model. An increase denotes that the edge is important while a decrease suggests otherwise.

Evaluating on the Test Set

The test samples are available inside data folder. To evaluate your model on the test set, please email us your predictions at swarna@cs.unc.edu.

The predictions should be generated in a tsv file with each line containing two tab-separated entries, first the predicted stance (support/counter) followed by the predicted graph in the same bracketed format as in the train and validation files. A sample prediction file in shown inside data folder.

For all latest results on ExplaGraphs, please refer to the leaderboard here.

Baseline Models

For training the stance prediction model, run the following script

bash model_scripts/train_stance_pred.sh

Note that this belongs to the rationalizing model family where the stance is predicted first and the predicted stance is conditioned on to generate the explanation graph. If you wish to work with the reasoning model family, append the generated linearized graph to the input by appropriately changing the src/utils_stance_pred.py file.

We also release the trained stance prediction model here. You can test the model on the validation split by running the following script

bash model_scripts/test_stance_pred.sh

For training and testing our structured graph generation model, refer to the README inside structured_model.

BART and T5-based graph generation models are coming soon!

Citation

@inproceedings{saha2021explagraphs,
  title={ExplaGraphs: An Explanation Graph Generation Task for Structured Commonsense Reasoning},
  author={Saha, Swarnadeep and Yadav, Prateek and Bauer, Lisa and Bansal, Mohit},
  booktitle={EMNLP},
  year={2021}
}