This Reads Like That

This is the code for the EMNLP 2023 publication: This Reads Like That: Deep Learning for Interpretable Natural Language Processing

Introduction

Prototype learning, a popular machine learning method designed for inherently interpretable decisions, leverages similarities to learned prototypes for classifying new data. While it is mainly applied in computer vision, in this work, we build upon prior research and further explore the extension of prototypical networks to natural language processing. We introduce a learned weighted similarity measure that enhances the similarity computation by focusing on informative dimensions of pre-trained sentence embeddings. Additionally, we propose a post-hoc explainability mechanism that extracts prediction-relevant words from both the prototype and input sentences. Finally, we empirically demonstrate that our proposed method not only improves predictive performance on the AG News and RT Polarity datasets over a previous prototype-based approach by Friedrich et al.(2022) but also enhances the faithfulness of explanations compared to rational-based work by Lei et al. (2016).

Setup

Installation

Clone this repository.

$ git clone https://github.com/fanconic/this_reads_like_that
$ cd this_reads_like_that

I suggest to create a virtual environment and install the required packages.

$ conda create --name nlp_env pytorch torchvision cudatoolkit=10.1 --channel pytorch
$ conda activate nlp_env
$ conda install --file requirements.txt

Repository Structure

• train.py: Main training loop
• rationales_training.py: Trains the model on the human-annotated rationale dataset and computes the faithfullness
• config.yaml: Config yaml file, which has all the experimental settings.
• run.sh: Runs the training script, including the conda environment. The config.yaml file is copied to experiment_configs
• gpu_experiments.sh: Runs the training script on a SLURM cluster, using GPUs. The output file is saved in experiment_outputs

Source Code Directory Tree

.
├── src                 # Source code            
    ├── layers              # Single Neural Network layers
    ├── models              # Neural Network Models
    ├── active              # Folder with functions for active learning
    ├── data                # Folder with data processing parts and datasets
        ├── embedding  
            ├── AG_NEWS         # precomputed embeddings of the AG NEWS dataset (needs to be downloaded and incerted from https://polybox.ethz.ch/index.php/s/S89h02V7AWDTlmw)
            ├── movies          # precomputed embeddings of the human-annotated movie reviews
            └── rt-polarity     # precomputed embeddings of the normal movie reviews
        ├── movies          # text data of the human-annotated movie reviews
        └── rt-polarity     # text data of the normal movie reviews   
    └── utils               # Useful functions, such as metrics, losses, etc
├── explanations        # Creates a CSV file with 50 test samples, and their according prototypes for interpretability
├── saved_models        # Saves the weights of the best neural networks
├── experiment_configs  # All the various configuration files for the experiments
└── experiment_outputs  # All outputs files of the experiments        

How to train

On your local computer:

bash run.sh <your_experiment_name>

On a SLURM cluster:

bash gpu_experiments.sh <your_experiment_name>

Reproduce the experiments

In order to reproduce the experiments with the ProtoTrex, you can download the embeddings from the various models and datasets from here https://polybox.ethz.ch/index.php/s/S89h02V7AWDTlmw. Subsequently, you can rerun our experiments with the following commands, where <model> should be changed with the backbone transformer (bert, gpt2, mpnet, roberta):

Weighted Similarity Experiments

• AG News - Cosine Similarity:

python3 -u train.py --config_path experiment_configs/<model>/cosine_news.yaml

• AG News - Weighted Cosine Similarity:

python3 -u train.py --config_path experiment_configs/<model>/weighted_cosine_news.yaml

• AG News - L2 Similarity:

python3 -u train.py --config_path experiment_configs/<model>/l2_news.yaml

• AG News - Weighted L2 Similarity:

python3 -u train.py --config_path experiment_configs/<model>/weighted_l2_news.yaml

• Movie Reviews - Cosine Similarity:

python3 -u train.py --config_path experiment_configs/<model>/cosine_movies.yaml

• Movie Reviews - Weighted Cosine Similarity:

python3 -u train.py --config_path experiment_configs/<model>/weighted_cosine_movies.yaml

• Movie Reviews - L2 Similarity:

python3 -u train.py --config_path experiment_configs/<model>/l2_movies.yaml

• Movie Reviews - Weighted L2 Similarity:

python3 -u train.py --config_path experiment_configs/<model>/weighted_L2_movies.yaml

Interpretability vs. Performance Trade-Off

• AG News - non-interpretable:

python3 -u train.py --config_path experiment_configs/<model>/non_interpretable_news.yaml

• Movie Reviews - non-interpretable:

python3 -u train.py --config_path experiment_configs/<model>/non_interpretable_movies.yaml

Ablation Studies of Loss

• Full Loss:

python3 -u train.py --config_path experiment_configs/bert/full_loss.yaml

• Without Clustering Loss:

python3 -u train.py --config_path experiment_configs/bert/remove_clust_loss.yaml

• Without Separation Loss:

python3 -u train.py --config_path experiment_configs/bert/remove_sep_loss.yaml

• Without Distribution Loss:

python3 -u train.py --config_path experiment_configs/bert/remove_distr_loss.yaml

• Without Diversity Loss:

python3 -u train.py --config_path experiment_configs/bert/remove_divers_loss.yaml

• Without L1 Regularizer:

python3 -u train.py --config_path experiment_configs/bert/remove_l1_loss.yaml

Faithfulness

• Train on human-annotated rationale movie review dataset

python3 -u rationales_training.py --config experiment_configs/<model>/rationales.yaml

Citation

@inproceedings{fanconi-etal-2023-reads,
    title = "This Reads Like That: Deep Learning for Interpretable Natural Language Processing",
    author = "Fanconi, Claudio  and
      Vandenhirtz, Moritz  and
      Husmann, Severin  and
      Vogt, Julia",
    editor = "Bouamor, Houda  and
      Pino, Juan  and
      Bali, Kalika",
    booktitle = "Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing",
    month = dec,
    year = "2023",
    address = "Singapore",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2023.emnlp-main.869",
    pages = "14067--14076"
}

Contributors

• Claudio Fanconi - fanconic@ethz.ch
• Moritz Vandenhirtz - moritz.vandenhirtz@inf.ethz.ch
• Severin Husmann - shusmann@ethz.ch
• Julia Vogt - julia.vogt@inf.ethz.ch