A Hybrid Approach to Skeleton-based Translation
Tong Xiao†‡, Jingbo Zhu†‡, Chunliang Zhang†‡
† Northeastern University, Shenyang 110819, China
‡ Hangzhou YaTuo Company, 358 Wener Rd., Hangzhou 310012, China
{xiaotong,zhujingbo,zhangcl}@mail.neu.edu.cn
Abstract
In this paper we explicitly consider sen-
tence skeleton information for Machine
Translation (MT). The basic idea is that
we translate the key elements of the input
sentence using a skeleton translation mod-
el, and then cover the remain segments us-
ing a full translation model. We apply our
approach to a state-of-the-art phrase-based
system and demonstrate very promising
BLEU improvements and TER reductions
on the NIST Chinese-English MT evalua-
tion data.
1 Introduction
Current Statistical Machine Translation (SMT) ap-
proaches model the translation problem as a pro-
cess of generating a derivation of atomic transla-
tion units, assuming that every unit is drawn out
of the same model. The simplest of these is the
phrase-based approach (Och et al., 1999; Koehn
et al., 2003) which employs a global model to
process any sub-strings of the input sentence. In
this way, all we need is to increasingly translate
a sequence of source words each time until the
entire sentence is covered. Despite good result-
s in many tasks, such a method ignores the roles
of each source word and is somewhat differen-
t from the way used by translators. For exam-
ple, an important-first strategy is generally adopt-
ed in human translation - we translate the key ele-
ments/structures (or skeleton) of the sentence first,
and then translate the remaining parts. This es-
pecially makes sense for some languages, such as
Chinese, where complex structures are usually in-
volved.
Note that the source-language structural infor-
mation has been intensively investigated in recent
studies of syntactic translation models. Some of
them developed syntax-based models on complete
syntactic trees with Treebank annotations (Liu et
al., 2006; Huang et al., 2006; Zhang et al., 2008),
and others used source-language syntax as soft
constraints (Marton and Resnik, 2008; Chiang,
2010). However, these approaches suffer from
the same problem as the phrase-based counterpart
and use the single global model to handle differ-
ent translation units, no matter they are from the
skeleton of the input tree/sentence or other not-so-
important sub-structures.
In this paper we instead explicitly model the
translation problem with sentence skeleton infor-
mation. In particular,
• We develop a skeleton-based model which
divides translation into two sub-models: a
skeleton translation model (i.e., translating
the key elements) and a full translation model
(i.e., translating the remaining source words
and generating the complete translation).
• We develop a skeletal language model to de-
scribe the possibility of translation skeleton
and handle some of the long-distance word
dependencies.
• We apply the proposed model to Chinese-
English phrase-based MT and demonstrate
promising BLEU improvements and TER re-
ductions on the NIST evaluation data.
2 A Skeleton-based Approach to MT
2.1 Skeleton Identification
The first issue that arises is how to identify the
skeleton for a given source sentence. Many ways
are available. E.g., we can start with a full syntac-
tic tree and transform it into a simpler form (e.g.,
removing a sub-tree). Here we choose a simple
and straightforward method: a skeleton is obtained
by dropping all unimportant words in the origi-
nal sentence, while preserving the grammaticali-
ty. See the following for an example skeleton of a
Chinese sentence.