How Synchronous are Adjuncts in Translation Data?

The argument-adjunct distinction is central to most syntactic and semantic theories. As optional elements that refine (the meaning of) a phrase, adjuncts are important for recursive, compositional accounts of syntax, semantics and translation. In formal accounts of machine translation, adjuncts are often treated as modifiers applying synchronously in source and target derivations. But how well can the assumption of synchronous adjunction explain translation equivalence in actual parallel data? In this paper we present the first empirical study of translation equivalence of adjuncts on a variety of FrenchEnglish parallel corpora, while varying word alignments so we can gauge the effect of errors in them. We show that for proper measurement of the types of translation equivalence of adjuncts, we must work with non-contiguous, many-to-many relations, thereby amending the traditional Direct Correspondence Assumption. Our empirical results show that 70% of manually identified adjuncts have adjunct translation equivalents in training data, against roughly 50% for automatically identified adjuncts.