Skip to content

Instantly share code, notes, and snippets.

@kowey
Created April 13, 2015 14:24
Show Gist options
  • Save kowey/4c45c65952c927c4698a to your computer and use it in GitHub Desktop.
Save kowey/4c45c65952c927c4698a to your computer and use it in GitHub Desktop.
class LinkPack(namedtuple('LinkPack',
['labels',
'pairings',
'scores_ad',
'scores_l'])):
'''
Collection of candidate links
:param labels: list of labels (same length as the width of
scores_l)
:type labels: [string]
:param pairings: list of EDU pairs (length = number of samples)
:type pairings: [EDU]
:param scores_ad: directed attachment scores (length should
match that of pairings)
:type scores_ad: array(float)
:param scores_l: label attachment scores (width should be
number of labels; height should be number
of pairings)
:type scores_l: matrix(float)
'''
def take(self, indices):
'''
Return a subset of the links indicated by the list/array
of indices
'''
# pylint: disable=no-member
return LinkPack(labels=self.labels,
pairings=np.take(self.pairings, indices),
scores_ad=np.take(self.scores_ad, indices),
scores_l=np.take(self.scores_l, indices))
# pylint: enable=no-member
def simple_candidates(self):
'''
Translate the links into a list of (EDU, EDU, float, string)
quadruplets representing a combined probability and best label
for each EDU pair. This is often good enough for simplistic
decoders
'''
# attachment * best label prob
# pylint: disable=no-member
scores = self.scores_ad * np.amax(self.scores_l, axis=1)
lbls = np.argmax(self.scores_l, axis=1)
# pylint: enable=no-member
return [(pair[0], pair[1], score, self.labels[lbl])
for pair, score, lbl
in zip(self.pairings, scores, lbls)]
class Decoder(with_metaclass(ABCMeta, object)):
'''
A decoder is a function which given a probability distribution (see below)
and some control parameters, returns a sequence of predictions.
Most decoders only really return one prediction in practice, but some,
like the A* decoder might have able to return a ranked sequence of
the "N best" predictions it can find
We have a few informal types to consider here:
- a **link** (`(string, string, string)`) represents a link
between a pair of EDUs. The first two items are their
identifiers, and the third is the link label
- a **candidate link** (or candidate, to be short,
`(EDU, EDU, float, string)`)
is a link with a probability attached
- a **prediction** is morally a set (in practice a list) of links
- a **distribution** is morally a set of proposed links
'''
@abstractmethod
def decode(self, lpack):
'''
:type lpack: LinkPack
:rtype: [ [(string,string,string)] ]
'''
raise NotImplementedError
class PruningDecoder(with_metaclass(ABCMeta, Decoder)):
'''
A pruning decoder takes another decoder as input and does some
preprocessing on the candidate edges, removing some of them
before handing them off to its inner decoder
'''
def __init__(self, decoder):
self.decoder = decoder
def decode(self, lpack):
return self.decoder.decode(self.prune(lpack))
@abstractmethod
def prune(self, lpack):
'''
Trim a set of proposed links
:rtype: LinkPack
'''
raise NotImplementedError
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment