How Synchronous are Adjuncts in Translation Data?
Sophie Arnoult, Khalil Sima’an · 2014
The argument-adjunct distinction is cen-tral to most syntactic and semantic the-ories. As optional elements that refine (the meaning of) a phrase, adjuncts are important for recursive, compositional ac-counts of syntax, semantics and transla-tion. In formal accounts of machine trans-lation, adjuncts are often treated as modi-fiers applying synchronously in source and target derivations. But how well can the assumption of synchronous adjunction ex-plain translation equivalence in actual par-allel data? In this paper we present the first empirical study of translation equiva-lence of adjuncts on a variety of French-English parallel corpora, while varying word alignments so we can gauge the ef-fect of errors in them. We show that for proper measurement of the types of trans-lation 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 manu-ally identified adjuncts have adjunct trans-lation equivalents in training data, against roughly 50 % for automatically identified adjuncts. 1