[developers] Questions from Korea - Yong-hun Lee

[Francis Bond]

G'day,

I am CCing this to developers, with permission, in order keep the
information widely available.

> I send this e-mail to ask some questions on machine translation system using the LOGON system. Since I am also a member of the project, I nowadays study how the machine translation system works in the LOGON system, using the DELPH-IN wiki. However, I have many difficulties to understand the basic mechanism of the system. Accordingly, I decided to ask some questions to you. I am very sorry, but please give me the answers for the following questions.
>
> [1] Is it impossible to make a machine translation system without the LinGO Grammar Matrix? I know that the LinGO Grammar Matrix is helpful to the machine translation system. What I want to know is if it is possible to make a machine translation system without the elements in the LinGO Grammar Matrix.

It is possible to make an MT system without the LinGO Grammar Matrix.
You do not even have to have HPSG grammars on both sides.

> [2] How “Rephrase” is different from “Translation” in the basic algorithm? Here, “Rephrase” means the mechanism which can generate (from the sentence “John eats an apple”) “John ate an apple”, “An apple John eats”, “An apple John ate”, and so on. Is it possible to make a “Rephrase” system without “mt.lsp”, “transfer.tdl”, “trigger.mtr”, “semi.vpm”, and so on? Which file is necessary for “Rephrase”, which file is necessary for “Translation”, and which file is necessary for both?

Rephrase uses a small set of transfer (paraphrase) rules that go from
one language into the same language.  As far as I know, only the ERG
uses this, and I don't know any details of how it is set up or whether
it is supported.   Dan Flickinger may know more.

semi.vpm and trigger.mtr are used even by monolingual generation.
mt.lsp and transfer.tdl are used by the transfer machinery (and I
guess the "rephrase" engine).

> [3] If the answer for [1] is Yes, how can I make a toy machine translation system? For example, can I make a toy machine translation system based on the “g7sem” grammar which is included in the LKB system? That is, can I make a toy machine translation system from “g7sem” grammar to “g7sem-trans” grammar? In the “g7sem-trans” grammar, all the ‘_rel’ in the lexical items are changed into ‘_trans’. For example, the word “the” has ‘the_rel’ in “g7sem” grammar, whereas the same word has ‘the_trans’ in “g7sem-trans” grammar. My question is if I can make a toy machine translation system which maps ‘the_rel’ in “g7sem” grammar into ‘the_trans’ in “g7sem-trans” grammar? If possible, how can I do it? Likewise, can I make a toy “Rephrase” system based on the “g7sem” grammar? If possible, how can I do it?

I am afraid that I only know how to make a relatively heavy MT system
such as JaEn or the EnEn paraphrasing system.    Sanghoun and I (and
others) set up a prototype KoEn system using the KRG and ERG, but it
is still a little buggy.

> In fact, both KRG and ERG are too big for me to understand the basic algorithms of the “Rephrase” and ”Translation” system. If there is a toy grammar which implemented the basic algorithms, it will be very helpful (to me) to understand the basic algorithms of the “Rephrase” and ”Translation” system. If you have such one, please let me know.

I think the transfer systems produced by Emily Bender in her class may
be more light weight, perhaps she or one of her students could help
answer.

> I am very grateful for your reading my mail, and I want your comments to my questions.

Sorry that I can't answer all of your questions, I hope someone else can,

--
Francis Bond <http://www3.ntu.edu.sg/home/fcbond/>
Division of Linguistics and Multilingual Studies
Nanyang Technological University