ETD system

Electronic theses and dissertations repository

 

Tesi etd-05252015-122546


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