ETD

Archivio digitale delle tesi discusse presso l'Università di Pisa

Tesi etd-05252015-122546


Tipo di tesi
Tesi di dottorato di ricerca
Autore
MICELI BARONE, ANTONIO VALERIO
URN
etd-05252015-122546
Titolo
Syntax-based machine translation using dependency grammars and discriminative machine learning
Settore scientifico disciplinare
INF/01
Corso di studi
SCIENZE DI BASE "GALILEO GALILEI"
Relatori
tutor Prof. Attardi, Giuseppe
relatore Prof.ssa Simi, Maria
relatore Prof. Pedreschi, Dino
Parole chiave
  • translation
  • dependency
  • syntax
  • learning
  • neural network
Data inizio appello
22/06/2015
Consultabilità
Completa
Riassunto
Machine translation underwent huge improvements since the groundbreaking
introduction of statistical methods in the early 2000s, going from very
domain-specific systems that still performed relatively poorly despite the
painstakingly crafting of thousands of ad-hoc rules, to general-purpose
systems automatically trained on large collections of bilingual texts which
manage to deliver understandable translations that convey the general
meaning of the original input.
These approaches however still perform quite below the level of human
translators, typically failing to convey detailed meaning and register, and
producing translations that, while readable, are often ungrammatical and
unidiomatic.
This quality gap, which is considerably large compared to most other
natural language processing tasks, has been the focus of the research in
recent years, with the development of increasingly sophisticated models that
attempt to exploit the syntactical structure of human languages, leveraging
the technology of statistical parsers, as well as advanced machine learning
methods such as marging-based structured prediction algorithms and neural
networks.
The translation software itself became more complex in order to accommodate
for the sophistication of these advanced models: the main translation
engine (the decoder) is now often combined with a pre-processor which
reorders the words of the source sentences to a target language word order, or
with a post-processor that ranks and selects a translation according according
to fine model from a list of candidate translations generated by a coarse
model.
In this thesis we investigate the statistical machine translation problem
from various angles, focusing on translation from non-analytic languages
whose syntax is best described by fluid non-projective dependency grammars
rather than the relatively strict phrase-structure grammars or projectivedependency
grammars which are most commonly used in the literature.
We propose a framework for modeling word reordering phenomena
between language pairs as transitions on non-projective source dependency
parse graphs. We quantitatively characterize reordering phenomena for the
German-to-English language pair as captured by this framework, specifically
investigating the incidence and effects of the non-projectivity of source
syntax and the non-locality of word movement w.r.t. the graph structure.
We evaluated several variants of hand-coded pre-ordering rules in order to
assess the impact of these phenomena on translation quality.
We propose a class of dependency-based source pre-ordering approaches
that reorder sentences based on a flexible models trained by SVMs and and
several recurrent neural network architectures.
We also propose a class of translation reranking models, both syntax-free
and source dependency-based, which make use of a type of neural networks
known as graph echo state networks which is highly flexible and requires
extremely little training resources, overcoming one of the main limitations
of neural network models for natural language processing tasks.
File