Closed texttheater closed 4 years ago
Idea: generate labeled dependencies, include "needs to be flipped" information into label.
Idea: embed “corresponding token” and “head” information into co/{4 => 5} terms. Extract dependencies by traversing each token's category. For a non-inverting argument, set the head on the argument to the owner of the current token. For an inverting argument, change the current head to the head of the argument. For no argument, set the owner on the CO to the current head. With categories like (s:_\np)/(s:adj\np)
, do something more complicated. Here, change the current head to the head of the argument and set the head of the np argument to that as well (i.e., process two arguments in one go). Then, recurse with the remaining category (without the processed argument(s)).
Action plan:
Our dependency scheme does not handle them in English V1 sentences, so this needs to be a post-processing step.