Machine Translation of Persian Complex Predicates

Complex Verbal Predicates (CPs) in Persian consist of a light verb combined with another nonverbal constituent, such as a noun, adjective, or adverb. The challenge of these constructions for machine translation (MT) is that they display both lexical and phrasal properties. They behave as lexical items in that the meaning of the construction is often non-compositional, but they also show phrasal properties as they can participate in syntactic processes; for instance, the constituents can be separated in an input sentence by intervening elements. In Persian, this problem is particularly severe, since most verbs are formed as complex constructions. Thus in many cases, Persian CPs present a translation divergence since they are translated as a single word in the target language (1). Traditionally, Persian translation systems have treated CPs as lexical items that are explicitly listed in the lexicon as compounds. The problem with this approach is that translation fails as soon as the CP elements are separated, e.g., by a direct object or clitic, modifiers, or relative clauses (2). A slightly more robust representation of CPs within an automatic translation system consists of incorporating their structure inside a complex lexicon, which would treat the preverbal element as “subcategorized ” by the light verb, listing the necessary syntax and translatio