📄 Identifying Word Correspondences in Parallel Texts

Summary

Gale and Church address word-level correspondence in bilingual corpora. The task is to find which English words translate which French words inside sentence pairs that have already been aligned. They distinguish this from sentence alignment because word order can cross between languages.

Working with the Canadian Hansards, they avoid the memory-hungry and possibly unstable EM estimation used by Brown et al. Instead, they score candidate word pairs with φ², a statistic computed from a 2×2 contingency table of co-occurrence counts over aligned regions. Pairs are selected iteratively and conservatively, with significance tests based on variance estimates and hand checks after each round. The accepted pairs are then placed within each sentence pair by a dynamic programming matcher. This matcher favours order-preserving matches and may leave words unmatched.

On a sample of 800 sentences, about 61% of English words received a French match, and roughly 95% of those matches were correct. The work favours precision over recall, so the output can be trusted for downstream use. Examples include probabilistic transfer dictionaries, word-based bilingual concordances, and sense-tagged corpora.

Key points

• Sentence alignment uses the authors' length-based method, since text lengths in characters correlate at 0.991 with their translations. About 90% of English sentences map to exactly one French sentence.
• For house/chambre, φ² is 0.62 with t = 123, computed over N = 897,077 aligned regions. The near miss house/communes is also significant (t = 31), but φ² separates the two pairs where mutual information does not.
• The authors criticise EM on two grounds:
  • Memory: Brown et al. limited each vocabulary to 9000 words, which already gives 81 million parameters.
  • Robustness: for hear→bravo, Brown et al. reported over 99%, while Gale and Church estimate about 60%.
• Word-based concordances reveal that take is rendered by some form of prendre in only 23% of its occurrences. IBM's estimate is 28%.
• The matcher's scoring has several parts:
  • Fan-in: the log-probability of a match is −0.05, −0.34 or −0.43 when a French position could match 1, 2 or many English words.
  • Slope: a slope of 1 is best, with log-probability −0.46.
  • NULL option: leaving a word unmatched has a fixed cost of about −5.
  • Errors: almost all remaining errors involve function words.

Interesting ideas

• I1. Prefer φ² to mutual information. φ² uses the off-diagonal cells, where one word appears without the other, so it better separates true translations from collocates. This is useful for any lexicon induction task built on co-occurrence counts.
• I2. Treat a score as a noisy estimate. A pair is accepted only if its φ² is significantly higher than the competing pairs, judged by its variance. This guards against spurious associations among rare words in any statistical pairing problem.
• I3. Select greedily and iteratively. Each round takes at most one pair per word and then removes already-chosen pairs. This reduces confusion from collocations such as House of Commons, and the approach transfers to entity linking or terminology extraction.
• I4. Make abstention a priced option. Giving the matcher a fixed cost for NULL lets one parameter control the precision/recall trade-off. Any aligner or tagger that should avoid guessing could use the same device.
• I5. Use translations as free evaluation labels. French renderings such as banque versus banc can grade English word sense disambiguation systems without manual annotation. The same trick could apply to other language pairs, including Ukrainian–English corpora.