negative documents construction for graph retriever of hotpotQA fullwiki about learning_to_retrieve_reasoning_paths HOT 2 CLOSED

Hi, @dsl-light

Thank you for going through our code!

TF-IDF documents first, then the hyperlink negative ones?

This is totally correct, based on the logic of our code.

My question is: hyperlink negative docs are considered by appending docs of all_linked_paras_dic, but keys of all_linked_paras_dic are all TF-IDF retrieved titles, so the most important part, hyperlink negative doc of gold path, may not be included for training?

For this, let us explain the logic in detail.

1. Appending gold paragraph titles (only during the training phase)
   https://github.com/AkariAsai/learning_to_retrieve_reasoning_paths/blob/master/graph_retriever/utils.py#L495
   https://github.com/AkariAsai/learning_to_retrieve_reasoning_paths/blob/master/graph_retriever/utils.py#L502

2. Appending TF-IDF-based negative examples
   https://github.com/AkariAsai/learning_to_retrieve_reasoning_paths/blob/master/graph_retriever/utils.py#L502
   We can control how many TF-IDF-based negative examples we use for the model training, and also please refer to https://github.com/AkariAsai/learning_to_retrieve_reasoning_paths/blob/master/graph_retriever/utils.py#L502
   for the use of the --tfidf_limit option.

3. Appending hyperlink-based negative examples
   https://github.com/AkariAsai/learning_to_retrieve_reasoning_paths/blob/master/graph_retriever/utils.py#L526
   https://github.com/AkariAsai/learning_to_retrieve_reasoning_paths/blob/master/graph_retriever/utils.py#L540
   Here we add hyperlink-based negative examples, and we can see that the hyperlinked titles are used.
   l is a hyperlinked paragraph's title from a paragraph p_ (example.all_linked_paras_dic[p_]).

Let us know if you have further quesitons.

Thank you!

from learning_to_retrieve_reasoning_paths.

@hassyGo Very clear explanation！Thank you!

from learning_to_retrieve_reasoning_paths.