RFB evaluation

[bohJiang12]

This PR aims to target issue #51

Implementations

• evaluate_alt.py: the alternative version of evaluation script, which borrows ideas of evaluate.py for calculating IOU. Its main functions include:
  • Load both prediction results run by RFB and gold-standard csv data, and return them as the same format -- {GUID: frame_num: (role, filler)} so that it's convenient for calculating IOU
  • The Intersection Over Union (IOU) evaluation metric, but missed the Dice-Sorensen Coefficient (DSC) for now.
  • The utility of writing out results.txt

Edge cases

All edge cases suffered now are from csv string (i.e. annotations part), and they're

• NaN value: this missing value occurs in either role or filler fields. -> Solution: replace this missing value as the string 'nan' while loading prediction results, but drop such row having it in loading gold results
• Address, Name, Company's name including comma: this confused how to successfully split role and filler fields from a csv string based on the rule of exporting gold data csv in process.py -> Solution: use str.split(',', maxsplit=2) to force the string is split into three fields: ('', role, filler)

Discussion

• In particular, I implemented a function get_aligned_ann_of for supporting find aligned annotation based on a "source" annotation cross views. Appreciate any feedback.
• The possible improvements can be
  • adding DSC metric
  • cleaner code for loading data

[bohJiang12]

@keighrim I just added the report, please take a look at that. Let me know if there's anything that I can improve.