Learning¶
In what follows,
- A refers to the attachment task: given an edu pair, is there a link between the two edus in any direction?
- D refers to the direction task: given an edu pair with a link between them, is the link from the textually earlier edu to the later on or vice-versa?
- L refers to the labelling task: given a directed linked edu pair, what is the label on edges between them?
- We use a ‘.’ character to denote the grouping of the tasks into models, so for example, an ‘AD.L’ scheme is one in which we use one model for predicting attachement and directions together; and a separate model for predicting labels
AD.L scheme¶
In the AD.L scheme, we learn
- a binary attachment/direction model on all edu pairs — (e1, e2) and (e2, e1) are considered to be different pairs here
- a multiclass label model on only the edu pairs that are have an edge between them in the training data
See decoding for details on how these models are used on decoding time
Probabilities¶
In the general case both attachment and labelling scores are probabilities, and so the resulting score is also a probability; however, this is not always appropriate for all classifiers.
For example, see this blog post on the implications of using a hinge loss function as opposed to the proper loss. If you are using a non-probability-based learner, you should also set –non-prob-scores to false on decoding time
Sometimes classifiers may not naturally support probabilities but can provide conversion mechanisms to compute them from scores. These methods may come with various downsides (eg. be expensive to compute, and more worryingly, inconsistent with the scores), so it may be best to stick with non-prob decoding for them too. See the note in the scikit manual for details.
Developers’ note: if you are developing classifiers for attelo, and your classifier does not return probabilties, it should implement decision_function instead