📄 Inducing Multilingual Text Analysis Tools via Robust Projection across Aligned Corpora

Summary

Most of the world's languages have no annotated corpora, so they lack statistical taggers and analyzers. The authors bootstrap such tools from English. They run existing English analyzers, such as the Brill tagger, on the English side of a parallel corpus. They then carry the resulting labels over to the other language through automatic word alignments produced by EGYPT (IBM Model 3). The corpora include the Canadian and Hong Kong Hansards, a Czech-English Reader's Digest collection and several Bible versions.

Labels transferred this way are very noisy, even when the alignments are perfect. To cope with this, the paper introduces noise-tolerant training procedures that turn the imperfect projections into stand-alone monolingual tools: POS taggers, base noun-phrase bracketers, named-entity taggers and lemmatizers. The case studies cover French, Chinese, Czech and Spanish.

For French, the induced tagger reaches 96% core-tag accuracy, the bracketer exceeds 91% F-measure, and verb lemmatization is above 99% accurate. None of this used hand-annotated data in the target language, and all results clearly beat raw projection. The approach suggests a cheap way to reuse English resources for under-resourced languages.

Key points

• Data: roughly 2 million words per language were used for tagging and bracketing.
  • Corpora: English-French Canadian Hansards, English-Chinese Hong Kong Hansards, Czech-English Reader's Digest.
  • Bibles: French Douay-Rheims, Spanish Reina Valera, and the King James, New International and Revised Standard versions in English.
  • Alignment: all word alignment was done with EGYPT, using raw words (characters for Chinese) and no stemming or dictionaries.
• French POS results:
  • Direct projection of core tags scores 76% with automatic alignments and 85% with manually corrected ones.
  • A standard bigram tagger trained on the projections reaches 86% on the aligned data and 82% on unseen French.
  • The noise-robust model scores 96% on the aligned data and 94% on unseen text. Fully supervised training reaches 97–98%.
  • The test set is a separate 200K-word hand-tagged French corpus from Université de Montréal.
  • Overall, the robust training cuts error by 71% relative to raw projection.
• Low tag ambiguity: words rarely carry more than two core POS tags.
  • In the Brown Corpus, only 0.37% of tokens fall outside a word's two most frequent tags.
  • In French, the figure is 0.03%.
• Morphology:
  • Lemmatization works even when the French Bible is the only bridge text, evaluated on Hansards verbs.
  • Accuracy improves steadily as the aligned corpus grows from 120K to 12M words.
  • Over 60% of errors come from roots missing from the automatically built candidate list.
  • Czech results are weaker because of poor alignments, a small corpus and near-duplicate perfective/imperfective roots.

Interesting ideas

• I1. Separate the lexical and sequence models. The tagger trains P(tag|word) and the tag-sequence model independently, and each gets its own noise treatment. This is useful whenever training labels are noisy, because different model components fail in different ways.
• I2. Amplify the dominant label per word. Each word's most frequent projected tag is boosted, while lower-ranked tags are shrunk or zeroed, and the same step is repeated at the subtag level. This exploits the fact that words are rarely ambiguous, and it could clean up any weakly supervised lexicon, including Ukrainian ones.
• I3. Weight training sentences by trustworthiness. Each sentence counts according to its alignment score and how well its projected tags agree with first-pass lexical priors. Because alignment errors come in bursts, dropping or downweighting weak regions improves quality while losing little useful data. The same idea fits any automatically labeled corpus.
• I4. Mark phrasal alignments explicitly first. When one English word aligns to several foreign words, each target word first receives a position-indexed tag. A learned mapping later converts these to ordinary tags, instead of forcing a guess at projection time.
• I5. Evaluate on a task with agreed answers. Morphology is scored by whether the system picks the correct root, which lexicographers largely agree on, rather than by theory-dependent analyses. The authors also note that for retrieval-style uses, grouping inflections correctly matters more than the exact lemma label.