LERG (Local Explanation of Response Generation) is a unified approach to explain why a conditional text generation model will predict a text. For more details, please refer to the paper Local Explanation of Dialogue Response Generation, Neurips 2021.
LERG can be installed from PyPI
pip install lerg
- Download our used data and model from Drive
- To explain dialogue generation models via methods
LERG_S,LERG_L, and baselinesSHAP,LIME, run
python explain.py --explain_method LERG_S --model_dir path-to-your-model-dir --data_path path-to-your-data-in-json
- After run
explain.py, get the<time-stamp>of saved explanations that stored in format"%m%d%Y_%H%M%S", run
python eval.py --explain_method LERG_S --time_stamp time-stamp-in-specified-format --model_dir path-to-your-model-dir --data_path path-to-your-data-in-json [--plot]
- To explain dialogue generation models via baselines
attn(attention),grad(gradient),none(random), run
python eval.py --explain_method attn --time_stamp None --model_dir path-to-your-model-dir --data_path path-to-your-data-in-json [--plot]
You can find Jupyter notebook in examples/.
The overall idea is to first import a purtabation model and an explanation method, then you can use LERG as:
from lerg.perturbation_models import RandomPM
from lerg.RG_explainers import LERG_SHAP_log as LERG_S
PM = RandomPM()
perturb_f = PM.perturb_inputs
local_exp = LERG_S(<replace_with_your_model_forward_function>, input_str, output_str, perturb_f, your_tokenizer)
phi_set, phi_map, input_segments, output_segments = local_exp.get_local_exp()
Perplexity Changes after removal (PPLC_R): (Slide with removal rate)
| Method | 0.1 | 0.2 | 0.3 | 0.4 | 0.5 |
|---|---|---|---|---|---|
| Random | 1.0699 | 1.1528 | 1.2317 | 1.3395 | 1.4449 |
| Attention | 1.0307 | 1.0931 | 1.1803 | 1.3106 | 1.479 |
| Gradient | 1.0578 | 1.1898 | 1.3244 | 1.4751 | 1.6340 |
| LIME | 1.1342 | 1.3667 | 1.5278 | 1.6807 | 1.8180 |
| SHAP | 1.1348 | 1.4038 | 1.5890 | 1.7823 | 1.9441 |
| LERG_L | 1.1442 | 1.3945 | 1.5554 | 1.7216 | 1.8646 |
| LERG_S | 1.2006 | 1.5588 | 1.8038 | 2.0332 | 2.2148 |
Perplexity with additives (PPL_A): (Slide with additive rate)
| Method | 0.5 | 0.6 | 0.7 | 0.8 | 0.9 |
|---|---|---|---|---|---|
| Random | 13.1789 | 14.0605 | 15.3738 | 16.5017 | 17.8391 |
| Attention | 13.4672 | 14.1226 | 14.9678 | 15.7989 | 16.9803 |
| Gradient | 12.7876 | 13.3019 | 14.1291 | 15.0506 | 16.1721 |
| LIME | 12.3468 | 12.5363 | 13.0447 | 13.5929 | 14.3374 |
| SHAP | 12.1357 | 12.2112 | 12.5584 | 12.9923 | 13.6296 |
| LERG_L | 12.0076 | 12.1196 | 12.5866 | 13.1717 | 13.8904 |
| LERG_S | 11.5665 | 11.4350 | 11.5816 | 11.9123 | 12.4341 |
If you find LERG is helpful to your research, we would appreciate a citation of this paper.
@article{tuan2021local,
title={Local Explanation of Dialogue Response Generation},
author={Tuan, Yi-Lin and Pryor, Connor and Chen, Wenhu and Getoor, Lise and Wang, William Yang},
journal={Advances in Neural Information Processing Systems},
volume={34},
year={2021}
}