Thermostat is a large collection of NLP model explanations and accompanying analysis tools.
- Combines explainability methods from the
captumlibrary with Hugging Face'sdatasetsandtransformers. - Mitigates repetitive execution of common experiments in Explainable NLP and thus reduces the environmental impact and financial roadblocks.
- Increases comparability and replicability of research.
- Reduces the implementational burden.
This work is described in our paper accepted to EMNLP 2021 System Demonstrations :
Nils Feldhus, Robert Schwarzenberg, and Sebastian Mรถller.
Thermostat: A Large Collection of NLP Model Explanations and Analysis Tools. 2021.
arXiv pre-print available here: https://arxiv.org/abs/2108.13961
Installation
With pip
pip install thermostat-datasets
Usage
Downloading a dataset requires just two lines of code:
import thermostat
data = thermostat.load("imdb-bert-lig")
Thermostat datasets can be addressed and loaded with an identifier string that contains three basic coordinates: Dataset, Model, and Explainer. In this example, the dataset is IMDb (sentiment analysis of movie reviews), the model is a BERT model fine-tuned on the IMDb data, the explanations are generated using a (Layer) Integrated Gradients explainer.
data then contains the following columns/features:
attributions(the attributions for each token for each data point; type: List of floats)idx(the index of the instance in the dataset)input_ids(the token IDs of the original dataset; type: List of ints)label(the label of the original dataset; type: int)predictions(the class logits of the classifier/downstream model; type: List of floats)
This is the raw content stored in each of the instances of data:
If we print data, we get more info such as the actual names of the dataset, the explainer and the model:
print(data)
> IMDb dataset, BERT model, Layer Integrated Gradients explanations
> Explainer: LayerIntegratedGradients
> Model: textattack/bert-base-uncased-imdb
> Dataset: imdb
Indexing an instance
We can simply index the loaded dataset like a list:
import thermostat
instance = thermostat.load("imdb-bert-lig")[429]
Visualizing attributions as a heatmap
We can apply .render() to every instance to display a heatmap visualization generated by the displaCy library.
instance.render() # instance refers to the variable assigned in the last codebox
Get simple tuple-based heatmap
The explanation attribute stores a tuple-based heatmap with the token, the attribution, and the token index as elements.
print(instance.explanation) # instance refers to the variable assigned in the second to last codebox
> [('[CLS]', 0.0, 0),
('amazing', 2.3141794204711914, 1),
('movie', 0.06655970215797424, 2),
('.', -0.47832658886909485, 3),
('some', 0.15708176791667938, 4),
('of', -0.02931656688451767, 5),
('the', -0.08834744244813919, 6),
('script', -0.2660972774028778, 7),
('writing', -0.4021594822406769, 8),
('could', -0.19280624389648438, 9),
('have', -0.015477157197892666, 10),
('been', -0.21898044645786285, 11),
('better', -0.4095713794231415, 12),
...] # abbreviated
The heatmap attribute displays it as a pandas table:
print(instance.heatmap)
> token_index 0 1 2 3 4 5 \
token [CLS] i went and saw this
attribution 0 -0.117371 0.0849944 0.165192 0.0362542 -0.029687
text_field text text text text text text
token_index 6 7 8 9 10 11 \
token movie last night after being coaxed
attribution 0.533126 0.240222 0.171116 -0.0450005 -0.0103401 0.0166524
text_field text text text text text text
token_index 13 14 15 16 17 \
token to by a few friends
attribution 0.0269605 -0.0213463 0.00761083 0.0216749 0.0579834
text_field text text text text text
# abbreviated
Modifying the load function
thermostat.load() is a wrapper around datasets.load_dataset() and you can use any keyword arguments from load_dataset() in load(), too (except path, name and split which are reserved), e.g. if you want to use another cache directory, you can use the cache_dir argument in thermostat.load().
Explainers
| Name | captum implementation | Parameters |
|---|---|---|
Layer Gradient x Activation (lgxa) |
.attr.LayerGradientXActivation |
|
Layer Integrated Gradients (lig) |
.attr.LayerIntegratedGradients |
# samples = 25 |
LIME (lime) |
.attr.LimeBase |
# samples = 25, mask prob = 0.3 |
Occlusion (occ) |
.attr.Occlusion |
sliding window = 3 |
Shapley Value Sampling (svs) |
.attr.ShapleyValueSampling |
# samples = 25 |
Datasets + Models
IMDb
imdb is a sentiment analysis dataset with 2 classes (pos and neg). The available split is the test subset containing 25k examples.
Example configuration: imdb-xlnet-lig
| Name |
|
lgxa |
lig |
lime |
occ |
svs |
|---|---|---|---|---|---|---|
ALBERT (albert) |
textattack/albert-base-v2-imdb |
|
|
|
|
|
BERT (bert) |
textattack/bert-base-uncased-imdb |
|
|
|
|
|
ELECTRA (electra) |
monologg/electra-small-finetuned-imdb |
|
|
|
|
|
RoBERTa (roberta) |
textattack/roberta-base-imdb |
|
|
|
|
|
XLNet (xlnet) |
textattack/xlnet-base-cased-imdb |
|
|
|
|
|
MultiNLI
multi_nli is a textual entailment dataset. The available split is the validation_matched subset containing 9815 examples.
Example configuration: multi_nli-roberta-lime
| Name |
|
lgxa |
lig |
lime |
occ |
svs |
|---|---|---|---|---|---|---|
ALBERT (albert) |
prajjwal1/albert-base-v2-mnli |
|
|
|
|
|
BERT (bert) |
textattack/bert-base-uncased-MNLI |
|
|
|
|
|
ELECTRA (electra) |
howey/electra-base-mnli |
|
|
|
|
|
RoBERTa (roberta) |
textattack/roberta-base-MNLI |
|
|
|
|
|
XLNet (xlnet) |
textattack/xlnet-base-cased-MNLI |
|
|
|
|
|
XNLI
xnli is a textual entailment dataset. It provides the test set of MultiNLI through the "en" configuration. The fine-tuned models used here are the same as the MultiNLI ones. The available split is the test subset containing 5010 examples.
Example configuration: xnli-roberta-lime
| Name |
|
lgxa |
lig |
lime |
occ |
svs |
|---|---|---|---|---|---|---|
ALBERT (albert) |
prajjwal1/albert-base-v2-mnli |
|
|
|
|
|
BERT (bert) |
textattack/bert-base-uncased-MNLI |
|
|
|
|
|
ELECTRA (electra) |
howey/electra-base-mnli |
|
|
|
|
|
RoBERTa (roberta) |
textattack/roberta-base-MNLI |
|
|
|
|
|
XLNet (xlnet) |
textattack/xlnet-base-cased-MNLI |
|
|
|
|
|
AG News
ag_news is a news topic classification dataset. The available split is the test subset containing 7600 examples.
Example configuration: ag_news-albert-svs
| Name |
|
lgxa |
lig |
lime |
occ |
svs |
|---|---|---|---|---|---|---|
ALBERT (albert) |
textattack/albert-base-v2-ag-news |
|
|
|
|
|
BERT (bert) |
textattack/bert-base-uncased-ag-news |
|
|
|
|
|
RoBERTa (roberta) |
textattack/roberta-base-ag-news |
|
|
|
|
|
Contribute a dataset
New explanation datasets must follow the JSONL format and include the five fields attributions, idx, input_ids, label and predictions as described above in "Usage".
Please follow the instructions for writing a dataset loading script in the official docs of datasets.
Provide the additional Thermostat metadata via the list of builder configs (click here to see the Thermostat implementation of builder configs).
Necessary fields include...
name: The unique identifier string, e.g. including the three coordinates<DATASET>-<MODEL>-<EXPLAINER>dataset: The full name of the dataset, usually follows the naming convention indatasets, e.g."imdb"explainer: The full name of the explainer, usually follows the naming convention incaptum, e.g."LayerIntegratedGradients"model: The full name of the model, usually follows the naming convention intransformers, e.g."textattack/bert-base-uncased-imdb"label_column: The name of the column in the JSONL file that contains the label, usually"label"label_classes: The list of label names or classes, e.g.["entailment", "neutral", "contradiction"]for NLI datasetstext_column: Either a string (if there is only one text column) or a list of strings that identify the column in the JSONL file that contains the text(s), e.g."text"(IMDb) or["premise", "hypothesis"](NLI)description: Should at least state the full names of the three coordinates, can optionally include more info such as hyperparameter choicesdata_url: The URL to the data storage, e.g. a Google Drive link
plus features which you can copy from the codebox below:
features={"attributions": "attributions",
"predictions": "predictions",
"input_ids": "input_ids"}
While debugging, you can wrap your data with the Thermopack class and see if it correctly parses your data:
import thermostat
from datasets import load_dataset
data = load_dataset('your_dataset')
thermostat.Thermopack(data)
If you're successful, follow the official instructions for sharing a community provided dataset at the HuggingFace hub.
At first, all Thermostat contributions will have to be loaded via the code example above. Please notify us of existing explanation datasets by creating an Issue with the tag Contribution and a maintainer of this repository will add your dataset to the Thermostat configs s.t. it can be accessed by everyone via thermostat.load().
Cite Thermostat
@inproceedings{feldhus2021thermostat,
title={Thermostat: A Large Collection of NLP Model Explanations and Analysis Tools},
author={Nils Feldhus and Robert Schwarzenberg and Sebastian Mรถller},
year={2021},
editor = {Heike Adel and Shuming Shi},
booktitle = {Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing: System Demonstrations},
}
Disclaimer
We give no warranties for the correctness of the heatmaps or any other part of the data. This is evolving work and will be hot-patched continuously.
The Thermostat project follows the ACL and ACM Code of Ethics.
Acknowledgements
The majority of the codebase, especially regarding the combination of transformers and captum, stems from our other recent project Empirical Explainers.
4.3k Jan 01, 2023
35 Sep 19, 2022
96 Dec 21, 2022
3 Aug 15, 2022
1 Jan 21, 2022
60 Nov 18, 2022
124 Dec 29, 2022
114 Nov 27, 2022
39 Jul 10, 2022
18 Dec 05, 2022
19 Sep 02, 2022
34 Oct 02, 2022
6 Jan 24, 2022
118 Dec 29, 2022
196 Jan 05, 2023
7.3k Dec 26, 2022
69 Dec 26, 2022
103 Dec 28, 2022
37 Dec 22, 2022
48 Nov 16, 2022