X-SPELLS-V2 - Explaining Short Text Classification with Diverse Synthetic Exemplars and Counter-Exemplars

Orestis Lampridis, Aristotle University of Thessaloniki, Greece, lorestis@csd.auth.gr
Laura State, University of Pisa, Italy and Scuola Normale Superiore, Pisa, Italy, laura.state@di.unipi.it
Riccardo Guidotti, University of Pisa, Italy, riccardo.guidotti@unipi.it
Salvatore Ruggieri, University of Pisa, Italy, salvatore.ruggieri@unipi.it

We present XSPELLS, a model-agnostic local approach for explaining the decisions of black box models in classification of short texts. The explanations provided consist of a set of exemplar sentences and a set of counter-exemplar sentences. The former are examples classified by the black box with the same label as the text to explain. The latter are examples classified with a different label (a form of counter-factuals). Both are close in meaning to the text to explain, and both are meaningful sentences – albeit they are synthetically generated. XSPELLS generates neighbors of the text to explain in a latent space using Variational Autoencoders for encoding text and decoding latent instances. A decision tree is learned from randomly generated neighbors, and used to drive the selection of the exemplars and counter-exemplars. Moreover, diversity of counter-exemplars is modeled as an optimization problem, solved by a greedy algorithm with theoretical guarantee. We report experiments on three datasets showing that XSPELLS outperforms the well-known LIME method in terms of quality of explanations, fidelity, diversity, and usefulness, and that is comparable to it in terms of stability.

Comments

This repository includes the code of the paper [1], which is a follow-up of [2]. X-SPELLS-V2 extends and replaces X-SPELLS.

Here, we uploaded basic data as well as source code files for the SVAE and BVAE. Computations for the OVAE are the same but require a few substitutions for data loading. We also uploaded the version of lime we used.

OVAE: Optimus Variational Autoencoder

We used the framework introduced in the paper [3] that can be found (code and models) at https://github.com/ChunyuanLI/Optimus.

To get the code running, proceed as follows:

1. Downloand code and one pretrained model
2. Pull the docker container of the project
3. Link the container to the code/models provided by OPTIMUS

Scripts and more details on how to get the models running are provided on the website. Some adaptation is necessary, e.g. to set data paths in the scripts accordingly or to insert code lines on generating examples in the latent space (no conceptual changes). Specifically, the following scripts are needed:

• A script to fine-tune the model, as provided on the OPTIMUS webpage, and the corresponding sh script.
• A script to manipulate the latent space, as provided on the OPTIMUS webpage, and the corresponding sh script. This python script needs to be augmented by the generate_sentences function from our create_explanations.py file, in order to create the necessary set of sentences in the latent space.

Instructions

To install the conda environment necessary to try out xspells run the following command:

conda env create -f environment.yml

In the BB_train directory, you can find the .py files to train the black boxes for all the datasets. As an example, to train the DNN black box for the hate speech dataset, we need to run the following command in the BB_train dir:

python hate_tweets_DNN.py

Afterwards, the model is saved inside the models directory for later use.

Next up, we can train the VAEs, located in the VAE_train folder. To train the BVAE for the hate speech, the polarity, the liar or the question dataset, type the following command in the VAE_train dir:

python train_BVAE.py

Make sure to change the dataset you want to use from inside the code. You can also use the youtube dataset by running the youtube_train_vae.py.

Finally, you can use the create_explanations.py script to use the models created above and produce the explanations.

Licenses

• MIT License for the source code in the lstm_vae directory.
• BSD 2-Clause "Simplified" License for the source code in the lime directory.
• Apache-2.0 License for the rest of the project.

References

[1] O. Lampridis, L. State, R. Guidotti, S. Ruggieri. Explaining short text classification with diverse synthetic exemplars and counter-exemplars. Machine Learning Journal, to appear, 2022.

[2] O. Lampridis, R. Guidotti, S. Ruggieri. Explaining Sentiment Classification with Synthetic Exemplars and Counter-Exemplars. Discovery Science (DS 2020). 357-373. Vol. 12323 of LNCS, Springer, September 2020

[3] Li, Chunyuan and Gao, Xiang and Li, Yuan and Li, Xiujun and Peng, Baolin and Zhang, Yizhe and Gao, Jianfeng, Optimus: Organizing Sentences via Pre-trained Modeling of a Latent Space, EMNLP, 2020