8.2 KiB


This repository contains the official code of the paper:

Neuro-Symbolic Visual Dialog [PDF]

Adnen Abdessaied, Mihai Bace, Andreas Bulling
International Conferenc on Computational Linguistics (COLING), 2022 / Gyeongju, Republic of Korea 🇰🇷 📢 Oral Presentation 📢

If you find our code useful or use it in your own projects, please cite our paper:

  author = {Abdessaied, Adnen and Bâce, Mihai and Bulling, Andreas},
  title = {{Neuro-Symbolic Visual Dialog}},
  booktitle = {Proceedings of the 29th International Conference on Computational Linguistics (COLING)},
  year = {2022},
  pages = {192--217},
  month = {oct},
  year = {2022},
  address = {Gyeongju, Republic of Korea},
  publisher = {International Committee on Computational Linguistics},
  url = {},
  pages = "192--217",


We propose Neuro-Symbolic Visual Dialog (NSVD) —the first method to combine deep learning and symbolic program execution for multi-round visually-grounded reasoning. NSVD significantly outperforms existing purely-connectionist methods on two key challenges inherent to visual dialog: long-distance co-reference resolution as well as vanishing question-answering performance. We demonstrate the latter by proposing a more realistic and stricter evaluation scheme in which we use predicted answers for the full dialog history when calculating accuracy. We describe two variants of our model and show that using this new scheme, our best model achieves an accuracy of 99.72% on CLEVR-Dialog —a relative improvement of more than 10% over the state of the art —while only requiring a fraction of training data. Moreover, we demonstrate that our neuro-symbolic models have a higher mean first failure round, are more robust against incomplete dialog histories, and generalise better not only to dialogs that are up to three times longer than those seen during training but also to unseen question types and scenes.


missingOverview of our method NSVD.

missingOverview of concat and stack encoders.


  • PyTorch 1.3.1
  • Python 3.6
  • Ubuntu 18.04

Raw Data

Scene Data

We used CLEVR and Minecraft images in this project. The raw images have a large footprint and we won't upload them. However, we provide their json file as well as their derendedred versions:

  • Original clevr-dialog training and validation raw scenes:
  • Raw scenes we used in our experiments:
  • All derendered scenes:

Dialog Data

The dialog data we used can be found here
You can also create your own data using the script.



The derendered scenes do not need any further preprocessing and can be directly used with our neuro-symbolic executor.


To preprocess the dialogs, follow these steps:

cd preprocess_dialogs

For the stack encoder, execute

python --input_dialogs_json <path_to_raw_dialog_file> --input_vocab_json '' --output_vocab_json <path_where_to_save_the_vocab> --output_h5_file <path_of_the_output_file> --split <train/val/test> --mode stack

For the concat encoder, execute

python --input_dialogs_json <path_to_raw_dialog_file> --input_vocab_json '' --output_vocab_json <path_where_to_save_the_vocab> --output_h5_file <path_of_the_output_file> --split <train/val/test> --mode concat


First, change directory

cd ../prog_generator

Caption Program Parser

To train the caption parser, execute

python --mode train --run_dir <experiment_dir> --res_path <path_to_store_results> --dataPathTr <path_to_preprocessed_training_data> --dataPathVal <path_to_preprocessed_val_data> --dataPathTest <path_to_preprocessed_test_data> --vocab_path <path_where_to_save_the_vocab>

Question Program Parser

To train the question program parser with the stack encoder, execute

python --mode train --run_dir <experiment_dir> --text_log_dir <log_dir_path> --dataPathTr <path_to_preprocessed_training_data> --dataPathVal <path_to_preprocessed_val_data> --dataPathTest <path_to_preprocessed_test_data> --scenePath <path_to_derendered_scenes> --vocab_path <path_where_to_save_the_vocab> --encoder_type 2

To train the question program parser with the concat encoder, execute

python --mode train --run_dir <experiment_dir> --text_log_dir <log_dir_path> --dataPathTr <path_to_preprocessed_training_data> --dataPathVal <path_to_preprocessed_val_data> --dataPathTest <path_to_preprocessed_test_data> --scenePath <path_to_derendered_scenes> --vocab_path <path_where_to_save_the_vocab> --encoder_type 1



To evaluate using the Hist+GT scheme, execute

python --mode test_with_gt --run_dir <experiment_dir> --text_log_dir <log_dir_path> --dataPathTr <path_to_preprocessed_training_data> --dataPathVal <path_to_preprocessed_val_data> --dataPathTest <path_to_preprocessed_test_data> --scenePath <path_to_derendered_scenes> --vocab_path <path_where_to_save_the_vocab> --encoder_type <1/2> --questionNetPath <path_to_pretrained_question_parser> --captionNetPath <path_to_pretrained_caption_parser> --dialogLen <total_number_of_dialog_rounds> --last_n_rounds <number_of_last_rounds_to_considered_in_history>

To evaluate using the Hist+Pred scheme, execute

python --mode test_with_pred --run_dir <experiment_dir> --text_log_dir <log_dir_path> --dataPathTr <path_to_preprocessed_training_data> --dataPathVal <path_to_preprocessed_val_data> --dataPathTest <path_to_preprocessed_test_data> --scenePath <path_to_derendered_scenes> --vocab_path <path_where_to_save_the_vocab> --encoder_type <1/2> --questionNetPath <path_to_pretrained_question_parser> --captionNetPath <path_to_pretrained_caption_parser> --dialogLen <total_number_of_dialog_rounds> --last_n_rounds <number_of_last_rounds_to_considered_in_history>


We achieve new state-of-the-art performance on clevr-dialog.


Model Accurcy NFFR
MAC-CQ 97.34 0.92
+ CAA 97.87 0.94
+ MTM 97.58 0.92
HCN 75.88 0.34
NSVD-concat (Ours) 99.59 0.98
NSVD-stack (Ours) 99.72 0.99


Model Accurcy NFFR
MAC-CQ 41.10 0.15
+ CAA 89.39 0.75
+ MTM 70.39 0.46
HCN 74.42 0.32
NSVD-concat (Ours) 99.59 0.98
NSVD-stack (Ours) 99.72 0.99

We refer to our paper for more results and experiments.


We thank Ahmed Shah for his MAC-XXX implemetation, Junki Ohmura for his HCN implemantation, Jiayuan Mao for providing us with the minecraft images, and finally Satwik Kottur for his clevr-dialog codebase.


For any questions or enquiries, don't hesitate to contact the above contributor.