How Can Transduction Grammar Induction Improve Statistical Machine Translation

PhD Qualifying Examination

Title: "How Can Transduction Grammar Induction Improve Statistical Machine 


Miss Nedjma Ousidhoum


In this survey, we study the limitations and the advantages of different 
transduction grammar formalisms in statistical machine translation in terms of 
representation, biparsing and training algorithms. We also examine the 
possibility of improving the translation quality given the grammatical 
structure of the output language, which typically is English.

Word alignment, formally describing the bilingual correlations between input 
and output languages, is a crucial step in the training of a statistical 
machine translation system. Learning a better word alignment highly impacts the 
translation quality. It may involve different resources and can be achieved in 
several ways such as formalizing the right rules and constraints given two 
independent monolingual structures and enforcing a common bilingual one. One of 
the challenges when designing such a model is to find a balance between both 
the heterogeneity of the different structures of human languages and the 
restrictions imposed by a specific set of translation examples.

We thoroughly survey research work on inversion transduction grammars, a 
bilingual equivalent to monolingual context free grammars, which have been 
shown to enhance the word alignment quality due to their hierarchical and 
compositional structure. We also review some further constrained transduction 
grammars such as linear transduction grammars and bracketing inversion 
transduction grammars. We inspect problems related to the representation of 
these formalisms: the creation of their syntactic rules, the induction of their 
lexical rules from given examples of translated sentences in addition to their 
parsing algorithms.

Date:			Tuesday, 15 August 2017

Time:                  	2:00pm - 4:00pm

Venue:                  Room 2126A
                         Lift 19

Committee Members:	Prof. Dekai Wu (Supervisor)
 			Prof. Andrew Horner (Chairperson)
 			Prof. Fangzhen Lin
 			Prof. Pascale Fung (ECE)

**** ALL are Welcome ****