Hello, I'm working on word alignment, currently trying to replicate some results from Och & Ney (2003) using GIZA++, as a sanity check before using GIZA++ as a baseline for my own experiments.
The problem is that I am getting suspiciously poor results with GIZA++, and I am unable to figure out why. Below is a summary of what I did, if anyone wants more detailed information please ask. Regards, Robert Östling Data: WPT-03 version of the English-French Hansards corpus, using the first 128k sentences as training data, and the test/trial sets for evaluation. This is lower-cased before fed to GIZA++, and I keep the original tokenization. I did not filter out any sentence pairs, but GIZA++ only reports a handful of sentences discarded. GIZA++ chain (performed separately in both directions): plain2snt.out XXX XXX snt2cooc.out XXX XXX XXX mkcls -c50 -n3 -pXXX -VXXX.classes mkcls -c50 -n3 -pXXX -VXXX.classes GIZA++ -compactalignmentformat 1 -s XXX -t XXX -c XXX -CoocurrenceFile XXX -m1 5 -m2 0 -mh 5 -m3 3 -m4 10 -o XXX Then I read (source index, target index) pairs from XXX.A3.final. Results: Here I use 128k training sentences, and intersection symmetrization, although the trend is the same in other experiments. 1. Och & Ney (2003) report an AER of 6.3%. 2. My own implementation gets 6.1% (test), which is unsurprising and indicates that the test setup is comparable to Och & Ney. 3. GIZA++ gets 8.9% (test). _______________________________________________ Moses-support mailing list [email protected] http://mailman.mit.edu/mailman/listinfo/moses-support
