head-comp-syntactic-learner

How to run the learner

To run the learning experiment, run main.py. Currently, it runs a sanity test which means it runs every type of input on every type of learner (16 scenarios). The experiment has several parameters:

• pp - whether to use PPs in the input.
• cp - whether to use CPs in the input.
• coordination - whether to have coordination structures in the input (the word "and").
• input_size - how many input items to generate (default - 50).
• temperature - the initial temperature for the SA learner (default - 100).

It is also possible to run the function test_learner which only runs one scenario. Its parameters are the type of learner, the type of input and then almost the same parameters as sanity_test.

Resources

• The logs for each run of the learner are saved in the folder logs.
• When running the parser, a parse tree can be generated. It will be saved to the folder resources.

Other notes

• The blank grammar being used is KAYNE_GRAMMAR_WITH_EMPTY_DP_NO_CO in the file BlankGrammars.py.
• The possible neighbour functions are in the variable self.neighbour_functions in the file MinimalistGrammarAnnealer.py.
  • This variable is initialized in the function set_learner_type and further modified in the function random_neighbour (so changes must be made in both places).
• The learner uses a numbered version of the parser - NumberBottomUpParser.py.
  • The original parser is in BottomUpParser.py and might not work.
  • The numbered parser translates the grammar items from their textual version to their numbered version and then parses.

Credits

• Original codebase written by Tomer Avraham.
• GA base class written by Matan Abudy.