Does translation software use syntactic trees?

Question

I'm a multilingual, and since syntax structures vary between various languages, it seems intuitive to me that translation software use syntactic rules to translate between languages. But when I see translation from English to Japanese, I'm puzzled as simple syntax rules seem to be ignored in the process.

Also, the same word can have different meanings adjacent to others, and that causes issues during computing a translation, but wouldn't Syntactic trees help with that? In knowing which phrase a morpheme is in, and what morphemes are in other parts of the phrase, shouldn't it be easier to compute a more accurate translation?

Answer 1

1. Purely statistical MT systems don't use parse trees. However most of today's MT systems use hybrid architectures, i.e. there's at least a shallow syntax module or something alike.

2. Parse trees can't help with semantic ambiguity. Whatever depends on context is dealt with later in the phase called (pragmatic) interpretation. Syntax provides some hints but what's decisive are semantic rules and constraints. Correct disambiguation always depends on context and common knowledge.

Answer 2

I have not worked personally on machine translation, but as I recall, some have been built, or experimented, using syntactic trees, and tree-to-tree translation mechanisms (I no longer have the references in mind).

Trees are of course syntactic, rather than semantic, though they can be decorated and usually are. However they can help with semantic ambiguity resulting from a syntactic ambiguity when languages have similarities in some of their syntactic structure.

There are clean formal ways to extend tree structures so that they encode several trees when there is syntactic ambiguity. Then you can consider translating this structure to the (extended) tree representation for the target language, and thus simply preserve the ambiguity. I remember people discussing this, but I have no idea whether it was actually ever implemented. This could for example be used for ambiguous prepositional attachment. But there are probably other ways.