SCFG translation

From cdec Decoder

cdec supports translating input sentences or lattices using synchronous context-free grammars (SCFGs), as described in Chiang (2007).

Example

Example grammar (fields are separated by a triple pipe, the first one in the left-hand side of the rule, the second is the source RHS, the third is the target RHS, and the optional fourth one is a list of feature values associated with the rule):

[S] ||| [NP,1] [VP,2] ||| [1] [2] ||| LogProb=0.2
[NP] ||| ich ||| i ||| LogProb=-0.6
[NP] ||| ein [NN,1] ||| a [1] ||| LogProb=-0.5
[NN] ||| [JJ,1] haus ||| [1] house ||| LogProb=-0.1
[NN] ||| [JJ,1] haus ||| [1] shell ||| LogProb=-1.3
[JJ] ||| kleines ||| small ||| LogProb=-0.1 SomeFeature=1.0
[JJ] ||| kleines ||| little ||| LogProb=-0.9 SomeFeature=0.2
[JJ] ||| grosses ||| big ||| LogProb=-0.8 SomeFeature=0.2
[JJ] ||| grosses ||| large ||| LogProb=-0.2345
[VP] ||| [V,1] [NP,2] ||| [1] [2] ||| LogProb=0.1 SomeOtherFeature=1.0
[V] ||| sah ||| saw ||| LogProb=-0.4
[V] ||| fand ||| found ||| LogProb=-0.4

An example weights file for use with the above grammar:

LogProb 1.0
SomeFeature 0.1
SomeOtherFeature 0.2

Translation of ich sah ein kleines haus (with 10-best output):

cdec -f scfg -g foo.grammar -k 10 -w we
cdec v1.0 (c) 2009 by Chris Dyer
Reading SCFG grammar from foo.grammar
  12 rules read.
Reading weights from we
Loaded 3 feature weights
Reading input from STDIN
ich sah ein kleines haus

INPUT: ich sah ein kleines haus
  id = 0
  Goal category: [S]
    .....
  -LM forest (nodes/edges): 8/10
  -LM forest       (paths): 4
  -LM Viterbi: i saw a small house
  -LM Viterbi: -1.1
0 ||| i saw a small house ||| LogProb=-1.4;SomeFeature=1;SomeOtherFeature=1 ||| -1.1
0 ||| i saw a little house ||| LogProb=-2.2;SomeFeature=0.2;SomeOtherFeature=1 ||| -1.98
0 ||| i saw a small shell ||| LogProb=-2.6;SomeFeature=1;SomeOtherFeature=1 ||| -2.3
0 ||| i saw a little shell ||| LogProb=-3.4;SomeFeature=0.2;SomeOtherFeature=1 ||| -3.18
Translation: 0.001966 secs (1 calls)

The full search space can be seen in the visualization example here.