About TrainingInstance module

[zhangsanaixuexi]

In your code, I noticed that there are many definitions of current_XXX and previous_XXX in the TrainingInstance module of the KBQA_Runner. Py file, and I want to ask you what current_topic_entity and previous_index and Candidate_paths2Previous_index represent..
I would appreciate it if you could give me your advice.

[lanyunshi]

Hi,

As described in the paper, we keep a beam to search for the relation paths iteratively and we employ reinforcement learning to update the model. To calculate the final rewards, each iteration we will remember the probabilities of the current top-ranked relations and multiply with its successive relations. the settings of previous_index, Candidate_paths2Previous_index are to remember the previous relations of the current relations.

A good way to understand the code is to print out some parts and see what is happening for each iteration. Thanks :)

[zhangsanaixuexi]

Thank you very much for your answer.
I want to take KBQA as the direction of my graduation thesis, but I notice that most of the current cutting-edge research in KBQA is about English. Is it feasible to use your paper method for reference to some Chinese data sets? What aspects are worthy of attention?

[lanyunshi]

Sorry for the late reply!

For Chinese datasets, I paid less attention to them. But I do notice that there is a Chinese KBQA challenge, you can find related information at link https://www.biendata.xyz/competition/ccks_2020_7_4/