📄 Using Coreference Links to Improve Spanish-to-English Machine Translation

Summary

The paper asks whether document-level coreference information can improve Spanish-to-English statistical machine translation, especially for pronouns. Spanish is pro-drop, so the system must sometimes create English pronouns from null subjects. The guiding idea is that a good translation should preserve the source text's grouping of mentions into entities.

Source-side links come from the gold annotation of the AnCora-ES corpus, which serves only as test data. Target-side links are computed automatically with the Stanford Statistical Coreference Resolution system. The authors test two strategies:

• Re-ranking: target mentions are projected onto source mentions and compared with standard coreference metrics (MUC, B3, CEAF). The resulting similarity is used to rescore Moses translation hypotheses.
• Post-editing: because the top hypotheses rarely differ in their mentions, this method combines pairwise coreference scores with MT scores and revises each mention individually.

Post-editing clearly improved pronoun accuracy while BLEU stayed essentially unchanged. Re-ranking helped pronouns but lowered overall and noun translation quality. This is presented as the first proof-of-concept that coreference links can measurably benefit statistical MT.

Key points

• Resources: AnCora-ES with gold coreference (source side, test only), Stanford CoreNLP coreference and its pairwise classifier (target side), and a Moses phrase-based baseline.

• Sanity check: on about 3.5k words, coreference similarity followed translation quality. The ranking was human translation, then commercial neural MT (BLEU 49.4), then the baseline (BLEU 43.7).

• Evaluation used BLEU for overall quality, APT for pronouns and ANT for nouns. Post-editing raised APT significantly with no significant change in BLEU or ANT.

• Manual evaluation covered four documents, nouns and pronouns together:

  System	Wrong mentions	Total score
  Baseline	53	251
  Re-ranking	55	249
  Post-editing	21	308

• The largest gains were for third-person personal and possessive pronouns, plus some Spanish null pronouns.