📄 Morphological Analysis for Statistical Machine Translation

Summary

The paper addresses a core difficulty in statistical machine translation between languages whose morphology differs sharply, using Arabic-to-English as the test case. A single Arabic orthographic word often packs several English words, such as an article, a preposition and a noun. Rare inflected forms are also poorly estimated. The author first splits Arabic words into prefix(es), stem and suffix(es), then POS-tags both sides of the parallel corpus. An automatic procedure built on IBM Model 1 alignments then decides, for each affix in a given stem-tag context, whether to keep it as a separate token, attach it back to its stem, or remove it. The goal is to make the two languages structurally closer. For phrase-based models, a hand-written rule additionally thins out redundant Arabic determiners inside noun phrases. Across training sets from 3.5K to 3.3M sentence pairs, BLEU improves substantially for both IBM Model 1 and a phrase translation model. The gains are largest when data is scarce. This shows that tailoring segmentation to the target language beats both raw words and maximal splitting.

Key points

• Prerequisites are fine-grained segmentation (following Lee et al. 2003), an Arabic POS tagger with about 120 tags from the LDC Arabic Treebank, and an English tagger with about 55 Penn Treebank tags. Tags are assumed to be single, with no combined tags like DET+NOUN.
• The decision rule compares two distributions:
  • the top-3 English tags predicted for each Arabic tag, and
  • the English tags predicted for each affix+tag within a stem-tag context.
• Affixes with no overlap become merge/delete candidates. A candidate is deleted if its most likely alignment is NULL and merged otherwise.
• Evaluation used LDC Multiple Translation Arabic Part I (1,043 segments from AFP and Xinhua) with uncased BLEU and 4 references. Training data mixed news, Ummah and UN text in four sizes: 3.5K, 35K, 350K and 3.3M pairs.
• IBM Model 1 BLEU rose from 0.10/0.14/0.18/0.18 to 0.25/0.29/0.31/0.32 across the four data sizes.
• The phrase model (monotone decoding) rose from 0.17/0.24/0.32/0.36 to 0.24/0.29/0.36/0.39.
  • Automatic analysis alone reached 0.21/0.25/0.33/0.36.
  • The manual Al#-deletion rule therefore added the rest of the gain, but it did not help Model 1.

Interesting ideas

• I1. Let the other language decide segmentation granularity. Splitting morphemes is not treated as good in itself. Each affix is kept separate only if it reliably maps to its own word class on the English side. This could guide how far to segment Ukrainian case endings or clitic-like particles for MT or alignment.
• I2. Do the statistics over POS tags instead of words. Working with tag-to-tag translation probabilities makes the merge/delete decisions robust and generalizable even with small corpora. It is a cheap way to induce linguistic rules from a parallel corpus.
• I3. Use alignment to NULL as a deletion signal. When a morpheme most often aligns to nothing, the paper drops it rather than gluing it to the stem. This is a simple, interpretable criterion that could be reused for function words or affixes in any language pair.
• I4. Condition affix behavior on the stem's part of speech. The same suffix may need different treatment after a noun than after a verb, so decisions are made per stem-tag context. This avoids one-size-fits-all preprocessing.
• I5. Tune preprocessing to the model type and report per-component gains. The determiner rule helped phrase models but not Model 1. The paper also reports automatic-only scores separately from automatic-plus-manual scores. This is a good evaluation habit for isolating which preprocessing step actually helps which system.