MMCoQA

MMCoQA aims to answer users’ questions with multimodal knowledge sources via multi-turn conversations.

For more details check out our ACL2022 paper "MMCoQA: Conversational Question Answering over Text, Tables, and Images".

Dataset download

The MMCoQA dataset can be downloaded via this link.

Dataset format

In the dataset folder you will find the following file question and contexts files:

1. MMCoQA_train.text,MMCoQA_dev.text,MMCoQA_test.text - contains questions, evidence and answers, for train, dev and test set respectively. Each line is a json format, that contains one question, the gold question, the conversation history, alongside its answers.
2. qrels.txt - contains the question and its relevant evidence (one document, table, or image) label.
3. multimodalqa_final_dataset_pipeline_camera_ready_MMQA_images.jsonl - contains metadata (id, title, path) for each image. You could load one image according to its path from the final_dataset_images.zip data.
4. multimodalqa_final_dataset_pipeline_camera_ready_MMQA_tables.jsonl - Each line represents a single table. header provides the column names in the table. table_rows is a list of rows, and each row contains is a list of table cells. Each cell is provided with its text string and link (if the text could be clicked in Wikipedia page).
5. multimodalqa_final_dataset_pipeline_camera_ready_MMQA_texts.jsonl - contains metadata (id, title, text) for each document.

Reporduce the MAE(pretrain) in our paper

1. Download the MMCoQA dataset and uncompress the zip file (including the final_dataset_images.zip).

2. MAE(pretrain) use the pretrained text based CoQA checkpoint from the work ORConvQA to initialize the text encoder. Please download it (checkpoint-5917) from this link and move it into the retriever_checkpoint folder.

3. Directly run train_retriever.py by setting the corresponding file paths.

python3 train_retriever.py 
--train_file MMCoQA_train.txt --dev_file MMCoQA_dev.txt --test_file MMCoQA_test.txt \
--passages_file multimodalqa_final_dataset_pipeline_camera_ready_MMQA_texts.jsonl \
--multimodalqa_final_dataset_pipeline_camera_ready_MMQA_tables.jsonl \
--images_file multimodalqa_final_dataset_pipeline_camera_ready_MMQA_images.jsonl \
--images_path final_dataset_images \
--retrieve_checkpoint ./retriever_checkpoint/checkpoint-5917

This script will store the checkpoints under the retriever_release_test folder. For your convenience, we upload our results (the checkpoint 'checkpoint-5061') in this link.

4. Generate embedding of evidence via setting 'retrieve_checkpoint' as the corresponding checkpoint file under the retriever_release_test folder.

python3 train_retriever.py 
--gen_passage_rep True \
--retrieve_checkpoint ./retriever_release_test/checkpoint-5061 \
--train_file MMCoQA_train.txt --dev_file MMCoQA_dev.txt --test_file MMCoQA_test.txt \
--passages_file multimodalqa_final_dataset_pipeline_camera_ready_MMQA_texts.jsonl \
--multimodalqa_final_dataset_pipeline_camera_ready_MMQA_tables.jsonl \
--images_file multimodalqa_final_dataset_pipeline_camera_ready_MMQA_images.jsonl \
--images_path final_dataset_images \

This script will store the embeddings of docs in the ./retriever_release_test/dev_blocks.txt file. For your convenience, we upload the dev_blocks.txt generated via the 'checkpoint-5061' in this link.

5. Run the train_pipeline.py.

python3 train_pipeline.py 
--train_file MMCoQA_train.txt --dev_file MMCoQA_dev.txt --test_file MMCoQA_test.txt \
--passages_file multimodalqa_final_dataset_pipeline_camera_ready_MMQA_texts.jsonl \
--multimodalqa_final_dataset_pipeline_camera_ready_MMQA_tables.jsonl \
--images_file multimodalqa_final_dataset_pipeline_camera_ready_MMQA_images.jsonl \
--images_path final_dataset_images \
--gen_passage_rep_output ./retriever_release_test/dev_blocks.txt \
--retrieve_checkpoint ./retriever_release_test/checkpoint-5061 \

This script will train the pipeline (retriever and answer extraction components) and store the checkpoints in release_test folder. For your convenience, we upload our results 'checkpoint-12000' in this link.

You could load the checkpoint and test it via:

python3 train_pipeline.py 
--do_train False --do_eval False --do_test True --best_global_step 12000 \
--train_file MMCoQA_train.txt --dev_file MMCoQA_dev.txt --test_file MMCoQA_test.txt \
--passages_file multimodalqa_final_dataset_pipeline_camera_ready_MMQA_texts.jsonl \
--multimodalqa_final_dataset_pipeline_camera_ready_MMQA_tables.jsonl \
--images_file multimodalqa_final_dataset_pipeline_camera_ready_MMQA_images.jsonl \
--images_path final_dataset_images \
--gen_passage_rep_output ./retriever_release_test/dev_blocks.txt \
--retrieve_checkpoint ./retriever_release_test/checkpoint-5061 \

The checkpoint 'checkpoint-12000' could achieve the results in the paper:

Dev set: {"f1": 4.586325704965762, "EM": 0.020654044750430294, "retriever_ndcg": 0.07209415622067673, "retriever_recall": 0.42168674698795183}

Test set: {"f1": 3.584818194987687, "EM": 0.0288135593220339, "retriever_ndcg": 0.07655641068040793, "retriever_recall": 0.4271186440677966}

The file of the prints while running train_pipeline.py is in this link.

Environment Requirement

The code has been tested running under Python 3.8.8 The required packages are as follows:

• pytorch == 1.8.0
• pytrec-eval == 0.5
• faiss-gpu == 1.7.0
• transformers == 2.3.0
• tensorboard == 2.6.0
• tqdm == 4.59.0
• numpy == 1.19.2
• scipy == 1.6.2

Contact

If there is any problem, please email liyongqi0@gmail.com. Please do not hesitate to email me directly as I do not frequently check GitHub issues.