Syntax Augmented Machine Translation: Challenges in Search and Learning Effec...

Google Tech Talks
December, 17 2007

ABSTRACT

Ashish Venugopal - RESEARCH SCIENTIST

Probabilistic Synchronous Context Free Grammars hold significant promise for machine translation, modeling context sensitive translation and re-odering effects with simple hierarchical operations learned directly from parallel data. Source language sentences are transformed into target language sentences via intermediate nonterminal symbols, typically via bottom up chart parsing with these grammars.

Introducing an N-Gram language model into this search space introduces dependencies between consecutive chart items, making exact search computationally difficult. We present a two pass approach that is motivated by grammars which include a large number of nonterminal symbols. We evaluate this method against a state of the art single pass approach.

The motivation for this two pass approach comes from a desire to include a large number of nonterminal labels in the translation grammar. Initial results using labels from associated phrase structure parse trees are promising, but this data is often noisy and requires human data generation. We propose a novel method to discriminatively learn nonterminal labels towards directly improving translation quality.

Speaker: Ashish Venugopal

Category:

People & Blogs

Tags:

• google
• techtalks
• techtalk
• engedu
• talk
• talks
• googletechtalks
• education

[zenoszond]

Interesting, thank you. I use CSG to parse rna structure. rnaparse com

[AI2flesh]

: 884509627386359275033751967 943067599621731590401694134 434007629683591574337516791 197615733475195375920401694 343151239621353184932676605 800621596380716399501371459 954387507655892533875618750 354029981152863950711207613. The "Only God is Perfect and I learned that.. happy journey you will go hug your wife after the journey trust me. hehe syntaxes cannot be meassured.

[AI2flesh]

that happyness I will never trade away again for knowledge, looser knowledge where you are not in circled of the real cores of life only a flight to a dimension your friends don´t understand

[AI2flesh]

watch you ass over years you will c your own jedi mind tricks. i know I am a dumb to listen to, well that suits me so fine, have worked hard to get back to beeing a dumb. DNA was normalized again for my happyness.

[AI2flesh]

the only way to learn the machine that we are seperate, that we sometimes find dephts that were better unseen, I hope you jewsbgs know what you are doing 10 years from now to your own mind, but time is essence then some might learn.

[AI2flesh]

so boerring to watch all the tech shit, just deleted my old account with no bans and warnings cuz i found the right knew it was about Syntaxes in all given to one, to make him c the big and relax with love.. close youtube for sure.I vote for it for sure cuz we become wise robots with too much iron in blood. so fucking sick of it. that depht hope you all get chocked and understand.

[0ffh]

you might be interested in probabilistic generative models - know wake-sleep learning, or the helmholtz machine... just to think a bit outside the box, while preserving much of your idea. with good intuition you could surely turn that into something you can evaluate in your framework.

[Skorrigan]

MT is still just a dream. Funny thing. People can fly to the moon, decompose DNA and construct powerful computers but they still can't come up with a programme that would correctly translate texts...

[vmanojnair]

Should speak slow. Quite difficult to understand.