Combining the Expression and Hierarchy Neural Nets

The next move is to combine the hierarchy and expression parser neural nets. For this I make a hierarchy where "move" and "bought" are a type of "infix", "to" is a type of "preposition" and "a" and the are types of "articles.

The expression parser is then setup with the "infix", "preposition" and "article" types. Where the priorities are "infix" < "preposition" < "article". I made a new class called Chain that can be used to take the output of the hierarchy neural net and use that as the input into the expression neural net. The main change for that was to add code to specify which nodes are outputs from the hierarchy and which nodes are input into the expression neural net. That is pretty much all that I did.

The expression parser is trained with this

priorities = [
  ('preposition', 'article'),
  ('infix', 'preposition'),
  ('constant', 'infix'),
  ('0', 'constant')
]

The hierarchy neural net is trained with this

words = [
  ('constant', 'constant'),
  ('to', 'preposition'),
  ('from', 'preposition'),
  ('move', 'infix'),
  ('bought', 'infix'),
  ('a', 'article'),
  ('the', 'article'),

  # top off the loop
  ('article', 'article'),
  ('preposition', 'preposition'),
  ('infix', 'infix')
]

The setup is done like so

one_hot_spec = ['constant', 'preposition', 'infix', 'article', '0']

hierarchy_model = HierarchyModel(words, 'constant', one_hot_spec)
parser_model = ParserModel(priorities, 'constant', one_hot_spec)

chain_model = ChainModel(hierarchy_model, parser_model)
chain_model.train(session)

Here is some sample usage:

Enter an sentence if you dare: move t1 to b2 dfhj ahadshfa sdhfgfah I bought a car
Input Expression: ['move', 't1', 'to', 'b2', 'dfhj', 'ahadshfa', 'sdhfgfah', 'I', 'bought', 'a', 'car']
Output Expression: {'action': 'move', 'thing': 't1', 'to': 'b2'}
Output Expression: {'buyer': 'I', 'action': 'buy', 'thing': {'thing': 'car', 'determiner': 'a'}}

The source code is expression5.py, hierarchy3.py and chain1.py . This is invoked by running expression5.py.  The source code is here