📄 Improved Alignment Models for Statistical Machine Translation

Summary

Word-to-word alignment models, in which each source word maps to exactly one target word, cannot capture links between groups of words. This leads to errors with reordering and with compounds such as German nouns that correspond to several English words. The authors study two ways to relax this restriction in statistical MT, tested on the Verbmobil German-English task (limited-domain spoken dialogues, 6000-word vocabulary) on both transcribed text and speech recognizer output.

1. Single-word approach. A monotone HMM-style alignment model with a trigram language model and DP beam search. It is extended with one-to-many alignments learned by training in the reverse direction, and with limited reordering for the German verb group.
2. Alignment template approach. Translation works on two levels: phrases are aligned to phrases, and words are aligned within each phrase pair. The templates are defined over automatically learned bilingual word classes.

The template approach gave lower error rates on every metric and translated about 30 times faster. This supports the claim that modelling word groups explicitly matters for MT.

Key points

• The test set has 147 sentences (1,968 words) from Verbmobil. Results are reported as WER, PER and a human-judged SSER, each with and without domain-specific preprocessing.
• On text input with preprocessing, the alignment template approach reached WER 48.3 / PER 35.1 / SSER 27.2. The single-word approach reached 56.0 / 41.2 / 35.3.
• On speech input, SSER was about 50–52 for templates and 53–55 for the single-word model. The recognizer's own word error rate was 31%, so the roughly 20% drop in translation quality was expected.
• Average time per sentence on an Alpha workstation was about 1 s for alignment templates and about 30 s for the single-word search.
• The single-word model allows only monotone jumps of width 0, 1 or 2. The quasi-monotone search permits reordering only for words of the German verb group.

Interesting ideas

• I1. Vocabulary extension from reverse-direction training. The authors train the model in the opposite direction (English to German) and collect adjacent English word sequences that align to a single German word. These sequences are added as new English vocabulary entries. This is a cheap way to obtain multiword units for compound-rich languages, for example Ukrainian–English or German-like morphology.
• I2. Templates over bilingual word classes instead of words. A phrase pattern learned from one town name or one number can then generalize to all members of the same class. This is useful wherever parallel data is sparse.
• I3. Splitting alignment into two levels. Phrase-to-phrase alignment handles word order, and a binary alignment matrix handles word links inside each phrase pair. Separating global and local structure this way could transfer to other sequence-pair tasks.
• I4. Using a position-independent error rate (PER) alongside WER. PER compares the bag of words in the output and the reference, ignoring order. Reporting both helps separate lexical errors from acceptable reordering, which matters for free-word-order languages.
• I5. Showing evaluators both systems' outputs at once. Human judges scored the two systems' translations side by side on a coarse 0 / 0.5 / 1 scale (correct / syntactically flawed / semantically wrong). This simple protocol is meant to make subjective comparisons more consistent and reproducible.