alignment-scripts

Scripts to preprocess training and test data for alignment experiments and to run and evaluate FastAlign and Mgiza.

Dependencies

• Python3
• MosesDecoder: Used during preprocessing
• Sentencepiece: Optional, used for subword splitting at the end of preprocessing
• FastAlign: Only used for FastAlign
• Mgiza: Only used for Mgiza

Usage Instructions

• Install all necessary dependencies
• Export install locations for dependencies: export {MOSES_DIR,FASTALIGN_DIR,MGIZA_DIR}=/foo/bar
• Make sure you set a reasonable default locale, e.g.: export LC_ALL=en_US.UTF-8
• Create folder for your test data: mkdir -p test
• Download Test Data for German-English and move it into the folder test
• Run preprocessing: ./preprocess/run.sh
• Run Fastalign: ./scripts/run_fast_align.sh
• Run Giza: ./scripts/run_giza.sh (This might take multiple days)

Results

All results are in percent in the format: AlignmentErrorRate (Precision/Recall)

German to English

Method	DeEn	EnDe	Grow-Diag	Grow-Diag-Final
FastAlign	28.4% (71.3%/71.8%)	32.0% (69.7%/66.4%)	27.0% (84.6%/64.1%)	27.7% (80.7%/65.5%)
Mgiza	21.0% (86.2%/72.8%)	23.1% (86.6%/69.0%)	21.4% (94.3%/67.2%)	20.6% (91.3%/70.2%)

Romanian to English

Method	RoEn	EnRo	Grow-Diag	Grow-Diag-Final
FastAlign	33.8% (71.8%/61.3%)	35.5% (70.6%/59.4%)	32.1% (85.1%/56.5%)	32.2% (81.4%/58.1%)
Mgiza	28.7% (82.7%/62.6%)	32.2% (79.5%/59.1%)	27.9% (94.0%/58.5%)	26.4% (90.9%/61.8%)

English to French

Method	EnFr	FrEn	Grow-Diag	Grow-Diag-Final
FastAlign	16.4% (80.0%/90.1%)	15.9% (81.3%/88.7%)	10.5% (90.8%/87.8%)	12.1% (87.7%/88.3%)
Mgiza	8.0% (91.4%/92.9%)	9.8% (91.6%/88.3%)	5.9% (97.5%/89.7%)	6.2% (95.5%/91.6%)

Known Issues

• Does not work on MacOs
• Tokenization of the Canadian Hansards seems to be off when accents are present in the English text: Ms. H é l è ne Alarie, Mr. Andr é Harvey :, Mr. R é al M é nard