Controlling Hallucinations at Word Level in Data-to-Text Generation

Code for our paper Controlling Hallucinations at Word Level in Data-to-Text Generation (Clément Rebuffel, Marco Roberti, Laure Soulier, Geoffrey Scoutheeten, Rossella Cancelliere, Patrick Gallinari).

With this code, you train a data-to-text encoder-decoder, whose decoder is multi-branch: during decoding, it weights several RNNs to compute probability of next token.

1. During training, weights are conditioned based on the probability of the target token to be a divergence from the table.
2. During inference, weights are fixed to only use RNNs which where associated to non-diverging tokens during training.

We also provide preprocessing scripts that work on WikiBio and ToTTo; as well as evaluation scripts for all metrics reported in the paper.

Note that most command are given for WikiBio, but will also work for ToTTo. If something is not straight forward, please open an issue or create a discussion.

Requirements

You can run this code using python 3.8. Using conda, you can create an env with:

conda create --name dtt python=3.8

Simply activate the env: conda activate dtt

You can install all needed packages using the following command:

pip3 install -r requirements.txt

Dataset: download & format

For all dataset related operations, please refer to data/README.md

We provide already an processed file for the WikiBio dataset: You can download a file where each line is a scored token, and examples are separated by empty lines

wget https://marcoroberti.cloudtb.online/wikibio_train_h.txt
mv download data/

This file can be processed into OpenNMT-readable file with:

python3 data/format_weights.py --orig data/download --dest train_weights.txt --strategy thresholds --thresholds 0.4 --normalize --weight_regularization 1 --eos_weights 1 0

This will use a fixed weight for the fluency factor (called in the script weight_regularization) and will give token to the content branch if they are scored below 0.4, else to the hallucination branch. --normalize means weights are normalized by their sum (so that they sum to 1) and --eos_weights 1 0 means that the End-of-Sequence token will be the responsability of the content branch.

Note that if you want to train a model on a small version of the dataset for practical reasons, you can create it with the following script:

python3 data/truncate_wikibio.py --folder small --max_size 1e4 --setname train test

Training

First things first, we compartmentalize experiments.

python3 create-experiment.py --dataset wikibio --name mbd

An experiment folder is now available at experiments/wikibio/mbd. You can move the previously created weights there:

mv train_weights.txt experiments/wikibio/mbd/train_weights.txt

We use the OpenNMT-py framework for training, included in onmt/. Our model has been added to our version of the repo. Training needs a preprocessing step:

mkdir experiments/wikibio/folder_with_dataset
python3 run_onmt.py --preprocess --config preprocess.cfg

Now that the preprocessing step is done, we can train a model using:

python3 run_onmt.py --train --config train_mbd.cfg

Please note that the model reported in our paper can be trained using the train_mbd.cfg config file.

Inference

The previous training step has saved checkpoints across time. In order to find the best check point, you can use the following command:

python3 batch_translate.py --dataset wikibio --setname valid --experiment small --bsz 64 --bms 10 --blk 0 --gpu 0 --weights 0.5 0.4 0.1

This will generate text for every saved checkpoints with a batch size 64 and a beam size 10.

Evaluation

BLEU and PARENT

We evaluate using the BLEU classic metric and PARENT, which is more suitable for the table-to-text generation task. Those n-gram based metrics can be computed as follows:

python3 compute_ngram_metrics.py data/wikibio/test_tables.jl data/wikibio/test_output.txt $OUTPUT_FILE

Hallucination rate

We can easily generate token-level hallucination score file in the same way we did for the training data (see the data/README.md file for details). Once obtained the $OUTPUT_SCORES file, we compute the mean of the sentence-level hallucination rates, considering as hallucinated all tokens that are above a given threshold:

python3 data/avg_hallucination_rate.py $OUTPUT_SCORES

Readability tests

We report the Flesch readability test. However note, that several classic readability tests can be performed :

• Kincaid
• ARI
• Coleman-Liau
• Flesch Index
• Fog Index
• Lix
• SMOG-Grading

Such values are computed using the GNU diction/style tool as follows:

cat $OUTPUT_FILE | sed -e 's/-lrb-/(/' -e 's/-rrb-/)/' -e 's/--/-/' -e "s/''/\"/" -e 's/``/"/' -e 's/./\u&/' | style