About implementation details, some questions:

[xmanners]

1. About Question reasoning type selection:

Notice that Question reasong type is determined before excuting graph segmentation, but seems like both the code & paper don't mention how to determine the type of complex question. According to the Experiment section, I wonder if the the question type was decided like method in DECOMPRC, one by a trained scorer and another is by a joint pipeline approach (Bert with classification head)?

2. About the QA reader in Experiment:

In Experiment section, the reader used two settings, one of them is a finetuned RoBERTa, and I don't find the description of another plain reader setting, so what's the other one?

3. Code running error?

When running the Bridge type question decompostion, an error occurs that is
assert list(set(verb_list).intersection(set(noun_list))) == []
when using T5-based stog & gtos model. Does it mean QDAMR not suitable for all kinds of multi-hop question in HotpotQA? If not, how to solve this assert error?

4. Sub-answer replacement

In Bridge-type question decomposition, after getting first sub-answer, it should replace the second-unknown part in the second sub-question, but after running the code, I found that in many cases, the selected second-unknown wasn't exist in the second-hop question, so I want to know to deal with this in your way.

Sorry to bother you & thank you very much for reading my Issue. Looking forward for your reply!

[xmanners]

Lastly there is one more question, in amrlib, seems like Spring-based gtos model doesn't exist but Experiment result contains it, where can I find that model?