Tutorial: Triggers #

However, due to possible combinatorial or computational constraints, or simply because a more controlled extraction process is wanted, the linguistic features included in the set of possible predictors are limited to a specific search space.

The feature space is defined in the YAML file (see Extraction)

Feature space #

For each node in the scope, it is possible to decide whether to include or exclude the features of their linear and dependency context.

E.g. Given the following scope with two nodes bound (N and A)

N[upos=NOUN]; A[upos=ADJ]; N-[amod]->A

For each of these nodes, we can add or exclude any feature encoded in the parent, previous, next, and the child nodes.

Note:

• It’s essential to remove features that are already defined in the scope and conclusion. Otherwise, the classifier will exploit this leaked information to make predictions, leading to an uninformative prediction.
• A node can have several children, and we choose to encode their features as sets. Therefore, the inclusion of a child’s feature should be interpreted as existential statement: there exists at least one child of X with feature A.
• Features are not included twice. For example, if the parent node of Y is already defined in the scope, it will not be added again.

In practice, the selection of features within the search space is made through the YAML file.

Check the included features #

The script check_features.py can be used to check the included features, helping to avoid any leaked information or undesirable features.