README.md

IBM Models 1 and 2

Python implementation of IBM Models 1 and 2. Tested using the a French-English parallel corpus from the HLT-NAACL 2003 Workshop.

We use various initializations for both model 1 and 2:

• Random initialization
• Uniform initialization
• Initialization of IBM model 2 with t from IBM model 1

We also added improvements to IBM model 1:

• Smoothing
• N-Null words

Training

Train the model by executing.

python src/main.py

Be aware we will save intermediate states of the model which may result in large data files (400MB+).

Evaluation

The python code will dump .eval files which can be used with the provided perl script: data/test/eval/wa_eval_align.pl. Or use bash eval.sh to execute all evaluations at once.