Future possibilities for MOCHEG

[f-tes]

Hi Barry Menglong Yao

I am a student working on a project on Fact-Checking. We really look up to your project and we think it's a really good model that we would like to reference in our work. We really admire all the good work you have done with Mocheg! Thank you so much for sharing it!

We have a few questions we would like to ask and we would really appreciate your guidance.

We were wondering 1) if it's possible for Mocheg to use news APIs such as Google News and NewsAPI instead of MSMARCO dataset for the module on evidence retrieval? If so, how would the code be different ?

2. Why does Mocheg train the evidence retrieval model ? Would the model still work if we do without it ?

Thank you so much!

[Barry-Menglong-Yao]

1. Do you mean that you want to change the document corpus from which we retrieve relevant documents? In other words, you want to retrieve relevant news based on Google News API, right? Maybe you can first check whether Google News API has already supported retrieving relevant news based on some keywords. I think you may not need to run our retrieval module to achieve that goal. However, if you have another news document corpus, you can adapt the logic in the line "corpus=data_dealer.load_corpus(args.test_data_folder, corpus_max_size)" in "retrieval/eval/eval_msmarco_mocheg.py" to load your specific corpus for retrieval.
2. Yes. It will still work without training. You can use a pre-trained evidence retrieval model. For example, you can do the inference with "python retrieve_train.py --mode=test --train_config=BI_ENCODER --model_name=all-mpnet-base-v2" and "python retrieve_train.py --mode=test --train_config=IMAGE_MODEL --model_name=clip-ViT-B-16", where "all-mpnet-base-v2" and "clip-ViT-B-16" are two pre-trained model from https://www.sbert.net/docs/pretrained_models.html.