A Pathbased Transfer Model for Machine Translation

1
A Path-based Transfer Model for Machine
Translation

• Dekang Lin
• presented by Joshua Johanson

2
Training

• Get a parallel corpus
• Source language is in dependency trees
• The text is word-aligned
• Extract the paths from dependency trees
• Learn translation rules from the paths using word
  alignment

3
Translation

• Parse the sentence into dependency trees
• Extract Paths
• Merge the paths
• Choose the transfer rules that give the highest
  probability
• Output the resulting sentence

4
What is a Dependency Tree?

• A dependency tree shows the relationship between
  the words of a sentence. Links are directed from
  the head to the modifier.

5
Comparing a Dependency Tree with a POS Tree
6
What is a path?
A simple path is a link of two nodes or two
links with an unassigned preposition.
7
Learning the Transfer Rules

• Extracts only paths with all words aligned
• A prepostion in the middle of a path is allowed
  to be aligned.
• Uses the word alignment to create the relative
  order of the paths. (there could be gaps)
• Learns the word alignment and the remapping.

8
Phrases

• Head span the word sequence aligned with the
  node n.
• Phrase span the word sequence from the lower
  bound of the head spans of all nodes in the
  subtree rooted at n to the upper bound of the
  same set of spans.
• All of these correspond to the index of the
  target language.

9

• Start with a simple path, lets say from Connect
  (H) to controller (M), where H aligns to Branchez
  (H) and M aligns to contrôleur (M). (A simple
  path can have a middle node with an unaligned
  preposition, like to.)

Connect
cables
power
both
to
the
controller
Branchez
les
deux
câbles
d
alimentation
sur
le
contrôleur
10

• Start with a simple path, lets say from Connect
  (H) to controller (M), where H aligns to Branchez
  (H) and M aligns to contrôleur (M). (A simple
  path can have a middle node with an unaligned
  preposition, like to.)
• Let S be the phrase span of a sibling of M (or
  head span of H) that is between H and M and
  closest to M. In this case it corresponds to
  câbles d alimentation (S).

Connect
cables
power
both
to
the
controller
Branchez
les
deux
câbles
d
alimentation
sur
le
contrôleur
11

• Start with a simple path, lets say from Connect
  (H) to controller (M), where H aligns to Branchez
  (H) and M aligns to contrôleur (M). (A simple
  path can have a middle node with an unaligned
  preposition, like to.)
• Let S be the phrase span of a sibling of M (or
  head span of H) that is between H and M and
  closest to M. In this case it corresponds to
  câbles d alimentation (S).
• The right hand side is the simple link in the
  original language

Connect
cables
power
both
to
the
controller
Branchez
les
deux
câbles
d
alimentation
sur
le
contrôleur
12

• Start with a simple path, lets say from Connect
  (H) to controller (M), where H aligns to Branchez
  (H) and M aligns to contrôleur (M). (A simple
  path can have a middle node with an unaligned
  preposition, like to.)
• Let S be the phrase span of a sibling of M (or
  head span of H) that is between H and M and
  closest to M. In this case it corresponds to
  câbles d alimentation (S).
• The right hand side is the simple link in the
  original language
• The left hand side is
• The link between H and M

Connect
cables
power
both
to
the
controller
Branchez
les
deux
câbles
d
alimentation
sur
le
contrôleur
13

• Start with a simple path, lets say from Connect
  (H) to controller (M), where H aligns to Branchez
  (H) and M aligns to contrôleur (M). (A simple
  path can have a middle node with an unaligned
  preposition, like to.)
• Let S be the phrase span of a sibling of M (or
  head span of H) that is between H and M and
  closest to M. In this case it corresponds to
  câbles d alimentation (S).
• The right hand side is the simple link in the
  original language
• The left hand side is
• The link between H and M
• A link between M and the nodes between S and the
  phrase span of M.

Connect
cables
power
both
to
the
controller
Branchez
les
deux
câbles
d
alimentation
sur
le
contrôleur
14

• Start with a simple path, lets say from Connect
  (H) to controller (M), where H aligns to Branchez
  (H) and M aligns to contrôleur (M). (A simple
  path can have a middle node with an unaligned
  preposition, like to.)
• Let S be the phrase span of a sibling of M (or
  head span of H) that is between H and M and
  closest to M. In this case it corresponds to
  câbles d alimentation (S).
• The right hand side is the simple link in the
  original language
• The left hand side is
• The link between H and M
• A link between M and the nodes between S and the
  phrase span of M.
• All unaligned word (like sur) will be leaf nodes.

Connect
cables
power
both
to
the
controller
Branchez
les
deux
câbles
d
alimentation
sur
le
contrôleur
15

• To align more complicated paths, just combine
  the translation of more simple paths
• This can create rules that are not paths

16
Divergences

• This will create dependency trees that are not
  consistent with the new language.
• In this case the translation will still produce
  the words in the correct order.

X
swim
across
Y
X
cruzar
nadando
Y
17
21 permutations
18
Generalize
19
Calculate Translation Probability

• Si is the path (Connect to controller)
• Ti is the tree fragment (Branchez sur contrôleur)
• c(Si) is the count of Si
• c(Ti,Si) is the count of both Ti and Si occuring
  together
• M is a constant

20
Translation

• Parse the sentence to obtain its dependency
  structure.

21
Translation

• Parse the sentence to obtain its dependency
  structure.
• Extract all the paths in the dependency tree and
  retrieve the translations of all the paths.

22
Translation

• Parse the sentence to obtain its dependency
  structure.
• Extract all the paths in the dependency tree and
  retrieve the translations of all the paths.
• Find rules that can be merged to cover the whole
  tree

23
Merging

• If two target nodes are mapped to the same source
  node, it gets merged.
• Merging will not create a loop
• We only have to worry about the unaligned words,
  which are leaf nodes and dont point to anything
• This new translation is a tree
• They are all connected and there arent any
  loops.

24
Node ordering

• If two nodes go on different sides of h, then go
  to the respective sides.
• deux câbles câbles existants
• deux câbles existants

25
Node ordering

• If they are on the same side as h in the target
  sentence, stay the same distance from h as in the
  source sentence.
• existing coaxial cables
• câbles coaxiaux existants

26
Node ordering

• If they are on the same side in the target
  sentence, but not the source sentence, use the
  word order of the original in the source sentence
• m1 h m2 (source)
• h m1 m2 (target)

27
Translation

• Parse the sentence to obtain its dependency
  structure.
• Extract all the paths in the dependency tree and
  retrieve the translations of all the paths.
• Find rules that can be merged to cover the whole
  tree
• Output the one with highest probability

28
Probability

• C is a set of paths covering S
• There can be overlap in C, but no path will
  completely contained in another in the final
  output.
• This is a direct translation (not noisy channel
  model)

29
Experiment

• Used English-French portion of 1999 European
  Parliament Proceedings.
• Used 1,755 sentences with 5-15 words out of
  116,889.
• Used Minipar to parse the sentences.
• Used ProAlign to align the words.

30
Results
31
What is different about this approach?

• Translations are based on a dependency tree in
  the source language
• Syntactically based
• There are fewer paths than subtrees (quadratic
  instead of exponential)
• Less sparse
• It automatically learns word order
• No need to know anything but syntax of target
  language