Conf: EACL 2014, Tutorial on Recent Advances in Dependency Parsing, Gothenburg, Sweden, 27 April 2014

Thierry Hamon hamon at LIMSI.FR
Tue Mar 11 21:14:18 UTC 2014

Date: Mon, 10 Mar 2014 14:14:43 +0100
From: peter ljunglöf <peter.ljunglof at>
Message-Id: <2F00416D-73BB-42E5-9DDF-C4FEE115CD29 at>


   EACL 2014 Tutorial: Recent Advances in Dependency Parsing
   Ryan McDonald, Joakim Nivre
   Gothenburg, Sweden, Sunday 27 April 2014

Syntactic parsing is a fundamental problem in natural language
processing which has been tackled using a wide variety of approaches.
In recent years, there has been a surge of interest in parsers that make
use of dependency structures, which offer a simple and transparent
encoding of predicate-argument structure and can be derived accurately
and efficiency using parsers trained on annotated corpora. Thanks to
their simplicity, transparency and efficiency, dependency parsers are in
widespread use for applications such as information extraction, question
answering, machine translation, language modeling, semantic role
labeling, and textual entailment.

This tutorial will focus on advances in dependency parsing that are not
covered in textbooks or previous tutorials, which means roughly work
from 2008 and onwards. However, in order to make the material accessible
to participants without a background in dependency parsing, we will
spend roughly the first quarter of the tutorial going over basic
concepts and techniques in the field, including the theoretical
foundations of dependency grammar and basic definitions of
representations, tasks, and evaluation metrics. After reviewing the
basic concepts, we will introduce the two dominant paradigms in early
work on data-driven dependency parsing -- global, exhaustive,
graph-based parsing and local, greedy, transition-based parsing -- and
review the contrastive error analysis presented in McDonald and Nivre
(2007), which highlighted the strengths and weaknesses of the two models
and set the challenge to improve both graph-based and transition-based
methods. This provides a basis for understanding many of the later
developments covered in the tutorial. The rest of the tutorial will be
divided into two main parts, covering advances in graph-based parsing
and related approaches, on the one hand, and advances based on
transition-based parsing, on the other. We will finish off with a
synthesizing conclusion and outlook for the future.

Research on graph-based dependency parsing in recent years has to a
large extent been driven by the wish to make efficient use of
higher-order models, thereby overcoming the limitations of strictly
local feature representations found in early models. As a consequence,
there has been developments towards specialized exact inference and
approximate inference methods, the latter especially for non-projective
parsing. In addition, there has been work on trying to find exact
dynamic programming solutions for restricted subsets of non-projective
structures, often referred to as mildly non-projective dependency trees.

Recent work on transition-based dependency parsing has focused on two
lines of research, often in combination. The first line has been
concerned with improving the search techniques through the use of beam
search, dynamic programming, and easy-first inference, thereby
overcoming the limitations of greedy left-to-right search. The second
line has been to improve the learning methods by moving to global
structured learning and or imitation learning with exploration, thereby
countering the negative effects of local classifier learning.  In
addition, there has been work on joint morphological and syntactic

Message diffuse par la liste Langage Naturel <LN at>
Informations, abonnement :
English version       : 
Archives                 :

La liste LN est parrainee par l'ATALA (Association pour le Traitement
Automatique des Langues)
Information et adhesion  :

ATALA décline toute responsabilité concernant le contenu des
messages diffusés sur la liste LN

More information about the Ln mailing list