[Mt-list] Machine Translation with Minimum Human Learning (MT/MHL)

[Yao Ziyuan]

I introduce a new MT concept, Machine Translation with Minimum Human
Learning (MT/MHL), which is a new approach to machine translation.
Before learning this concept, you must first learn another concept,
Automatic Code-Switching (ACS), which is a new approach to foreign
language learning. You can also find a formatted version of both
concepts at http://sites.google.com/site/yaoziyuan/ideas .

1. Automatic Code-Switching (ACS)

The computer automatically selects a few words in a user's native
language communication (such as a web page being viewed), and
supplements or even replaces them with their foreign language
counterparts, thus naturally building up his vocabulary. For example,
if a sentence
        他是一个好学生。
(Chinese for "He is a good student.") appears in a Chinese person's
Web browser, the computer can insert student after 学生 (optionally with
additional information such as student's pronunciation):
        他是一个好学生 (student)。
After several times of such teaching, the computer can directly
replace future occurrences of 学生 with student:
        他是一个好 student。
Ambiguous words such as the 看 (Chinese for "see", "look", "watch",
"read", etc.) in
        他在电视前看书。
(Chinese for "He is reading a book before the TV.") can also be
automatically handled by listing all context-possible translations:
        他在电视前看 (阅读: read; 观看: watch) 书。
Practice is also possible:
        他在电视前 [read? watch?] 书。
Because the computer would only teach and/or practice foreign language
elements at a small number of positions in the native language article
the user is viewing, the user wouldn't find it too intrusive.
Automatic code-switching can also teach grammatical knowledge in
similar ways.

2. Machine Translation with Minimum Human Learning (MT/MHL)

Before artificial intelligence reaches its fullest potential, machine
translation always faces unresolvable ambiguities. The good news is,
statistical MT such as Google Translate disambiguates content words
quite well in most cases, and syntactic ambiguity can largely be
"transferred" to the target language, without being resolved, if both
the source and the target language have common syntactic features. For
example, both English and French support prepositional phrases, so
        I passed the test with his help.
can be translated to French without determining whether "with his
help" modifies "passed" or "the test". The bad news is, syntactic
disambiguation usually can't be bypassed between a language pair like
English to Chinese, as in Chinese you must put "with his help" before
the constituent it modifies ("passed" or "the test"). Resolution of
syntactic ambiguity requires capabilities ranging from shallow rules
(e.g. "with help" should modify an action rather than an entity, and
if there are more than one action -- both "pass" and "test" can be
considered actions -- it should modify the verb -- "pass") to the most
sophisticated reasoning based on context or even information external
to the text (e.g. in
        Do you see the cat near the tree and the man?
what is the prepositional object of "near"? "The tree" or "the tree
and the man"?)

In light of automatic code-switching (ACS), an emerging technology
that promises to make foreign language learning efficient and
effortless, we can actually let an ACS system teach a human essential
syntactic knowledge (including key prepositions) of a foreign language
(or language family), so that machine translation can simply take care
of content word translation and leave syntactic puzzles "as is" to the
human. For example, when translating
        I passed the test with his help.
to Chinese, instead of trying to determine which word is modified by
"with his help", the MT system simply retains the original word order
and preposition ("with"), and just translates the content words:
        我 通过了 测试 with 他的 帮助。
which literally means "I PASSED THE-TEST with HIS HELP".

Best Regards,
Yao Ziyuan
http://sites.google.com/site/yaoziyuan/

_______________________________________________
Mt-list mailing list