Agreement Matters: Challenges of Translating into a Morphologically Rich Language, and the Advantages of a Syntax-Based System

Consider the following (simple) English sentences: “I drive a car.”, “I don’t know how to drive”, “I wash the car”, “I wash the floor”. Translating them to Hebrew using Google’s statistical MT system, yields: zipekna bdep ip` (I drive(masculine) a car); bedpl zr ei `l ip` (I don’t know(feminine) how to drive); ugex ip` zipeknd z` (I wash(masculine) the car); and dtvxd z` zthey ip` (I wash(feminine) the floor). While amusing and not quite politically correct, these are all arguably very good translations: without explicit gender marking, the translator can not know if the speaker is masculine or feminine, and he (she?) resorts to deciding based on her (his?) cultural knowledge. This does, however, highlight a class of problems which arise when attempting to translate from a morphologically clean language (e.g. English) into a morphologically rich one (e.g. Hebrew): many words in the target language are morphologically marked for gender and number, and the translator should be able to generate these markings correctly, based on little, elusive or sometimes no evidence in the source language. These issues are orthogonal to the data sparsity issues associated with highly inflected languages. Can current state-of-the-art statistical MT systems handle this? In what follows we present a few cases where the target language output should be morphologically marked for either gender or number, with varying amounts (and sources) of information available on the source language text, and discuss the suitability of current translation models to handle these phenomena. We show that correct handling of morphological agreement is beyond the reach of current systems as it requires better syntactic models, looking beyond a single sentence, and performing accurate anaphora resolution. However, while phrase-based models can not model even the simplest cases, syntax based models already posses most of the necessary machinery. While we demonstrate using English ⇒ Hebrew translations, similar issues will occur when translating into practically any morphologically rich language. Moreover, the issues discussed remain relevant also when the source language is also morphologically rich.