This is the code to learn alignments from interacting with wikiHow protocols.
The data folder includes the formatted protocols, the computed dependencies, and the computed semantic alignment betwen pairs of protocols.
The evaluation folder includes a human evaluation performed by an expert on a sample of the alignments obtained with 1500 training interactions.
The file aligner_agents.py includes the code to make a pair of agents interact and find an alignment. The experiment can be executed typing in a console
python aligner_agents.py [-t,-r]
In the experiment, the agents perform a training session of variable length, and then 100 test interactions. The three accepted parameters are the following:
-t : training interactions. Changes the amount of training interactions agents go through. Default is 100.
-r : Protocol set index. Must be between 0 and 4 (inclusive). Chooses one of the five randomly selected sets of training + test protocols. These sets are in the file protocols.py.
NOTE: we do not include the code to obtain the precision and recall of the alignment against the Oxford Dictionary because we do not have permision to make public the data in the dictionary.