# Efficient Beam Thresholding for Statistical Machine Translation

**ABSTRACT** Beam thresholding is a widely-used pruning approach in decoding algorithms of statistical machine translation. In this paper, we pro-pose two variations on the conventional beam thresholding, both of which speed up the de-coding without degrading BLEU score. The first variation is the dynamic beam threshold-ing, in which the beam threshold varies with the length of source sequences covered by hy-potheses. The second one incorporates a lan-guage model look-ahead probability into the beam thresholding so that the interaction be-tween a hypothesis and the contexts outside the hypothesis can be captured. Both thresh-olding methods achieve significant speed im-provements when used separately. By com-bining them together, we obtain a further speedup, which is comparable to that of the cube pruning approach (Chiang, 2007). Ex-periments also display that the dynamic beam thresholding can further improve the cube pruning.

**0**Bookmarks

**·**

**44**Views

- Citations (15)
- Cited In (0)

- [Show abstract] [Hide abstract]

**ABSTRACT:**In this paper, we describe an efficient A* search algorithm for statistical machine translation. In contrary to beamsearch or greedy approaches it is possible to guarantee the avoidance of search errors with A*. We develop various sophisticated admissible and almost admissible heuristic functions. Especially our newly developped method to perform a multi-pass A* search with an iteratively improved heuristic function allows us to translate even long sentences.07/2003; - SourceAvailable from: Deyi Xiong
##### Conference Paper: Maximum Entropy Based Phrase Reordering Model for Statistical Machine Translation.

[Show abstract] [Hide abstract]

**ABSTRACT:**We propose a novel reordering model for phrase-based statistical machine transla- tion (SMT) that uses a maximum entropy (MaxEnt) model to predicate reorderings of neighbor blocks (phrase pairs). The model provides content-dependent, hier- archical phrasal reordering with general- ization based on features automatically learned from a real-world bitext. We present an algorithm to extract all reorder- ing events of neighbor blocks from bilin- gual data. In our experiments on Chinese- to-English translation, this MaxEnt-based reordering model obtains significant im- provements in BLEU score on the NIST MT-05 and IWSLT-04 tasks.ACL 2006, 21st International Conference on Computational Linguistics and 44th Annual Meeting of the Association for Computational Linguistics, Proceedings of the Conference, Sydney, Australia, 17-21 July 2006; 01/2006 - SourceAvailable from: upenn.edu
##### Conference Paper: A Polynomial-Time Algorithm for Statistical Machine Translation.

[Show abstract] [Hide abstract]

**ABSTRACT:**We introduce a polynomial-time algorithm for statistical machine translation. This algorithm can be used in place of the expensive, slow best-first search strategies in current statistical translation architectures. The approach employs the stochastic bracketing transduction grammar (SBTG) model we recently introduced to replace earlier word alignment channel models, while retaining a bigram language model. The new algorithm in our experience yields major speed improvement with no significant loss of accuracy.34th Annual Meeting of the Association for Computational Linguistics, 24-27 June 1996, University of California, Santa Cruz, California, USA, Proceedings.; 01/1996

Page 1

Efficient Beam Thresholding for Statistical Machine Translation

Deyi Xiong, Min Zhang, Aiti Aw and Haizhou Li

Human Language Technology

Institute for Infocomm Research

1 Fusionopolis Way, #21-01 Connexis, Singapore 138632

{dyxiong, mzhang, aaiti, hli}@i2r.a-star.edu.sg

Abstract

Beam thresholding is a widely-used pruning

approach in decoding algorithms of statistical

machine translation. In this paper, we pro-

pose two variations on the conventional beam

thresholding, both of which speed up the de-

coding without degrading BLEU score. The

first variation is the dynamic beam threshold-

ing, in which the beam threshold varies with

the length of source sequences covered by hy-

potheses. The second one incorporates a lan-

guage model look-ahead probability into the

beam thresholding so that the interaction be-

tween a hypothesis and the contexts outside

the hypothesis can be captured. Both thresh-

olding methods achieve significant speed im-

provements when used separately. By com-

bining them together, we obtain a further

speedup, which is comparable to that of the

cube pruning approach (Chiang, 2007). Ex-

periments also display that the dynamic beam

thresholding can further improve the cube

pruning.

1 Introduction

Most statistical machine translation (SMT) mod-

els are of high complexity, especially when inter-

sected with an m-gram language model.

duce the search space, practical decoding algorithms

make use of pruning techniques. The most widely

used methods are beam thresholding (also known as

threshold pruning) and histogram pruning. The for-

mer retains only hypotheses with a probability larger

than a factor α of the currently best hypothesis. The

latter limits the number of retained hypotheses to

To re-

a maximum number b. In this paper, we examine

the beam thresholding and focus on two problems of

the traditional beam thresholding in SMT decoding.

We propose two efficient variations to address these

limitations. Our goal is to further reduce the search

space explored by the traditional beam thresholding

so that we can speed up the decoding.

The first problem is that the traditional beam

thresholding uses a fixed probability threshold (α)

throughout the whole decoding process. This is not

a good approach as the search space is not uniformly

distributed during decoding. We have found that a

dynamic beam thresholding, in which we adjust the

beam threshold according to the length of source se-

quence covered by hypotheses, can remove 60% of

hypotheses explored by the traditional beam thresh-

olding without the risk of decreasing BLEU (Pap-

ineni et al., 1996) score.

The second problem with the traditional applica-

tion of beam thresholding in SMT decoding is that

it only uses the probability estimated from inside a

hypothesis. It does not look outside the hypothesis:

the interaction of the current hypothesis with outside

contexts. Consider the case in which a particular hy-

pothesis within the current beam is to be expanded

further while after one or more expansion steps, the

sub-hypothesis falls outside the beam. The expan-

sion efforts are wasted if we do not generate n-best

lists. Therefore can we prune such a hypothesis be-

fore we expand it further? The key to the question is

to find another probability to measure the future in-

teraction of the current hypothesis with outside con-

texts so that we can threshold out bad hypotheses as

early as possible. We use a language model look-

Page 2

ahead probability to evaluate the future interaction.

We test the two proposed beam thresholding

methods on a CKY-style decoder with beam search,

which is developed for BTG-based SMT (Wu,

1996). The CKY-style decoder uses a bottom-up

CKY chart parsing algorithm. All hypotheses cover-

ing the same span of the source sentence are stored

in a cell of the chart. The decoder compares hy-

potheses to other hypotheses in the same cell and

thresholds out bad hypotheses.

sults show that two beam thresholding methods lead

to significant speedups over the traditional beam

thresholding, when used separately.

ing them together, we achieve a further improve-

ment of the decoding speed, which is comparable

to that of the cube pruning. When we apply our dy-

namic beam thresholding to the cube pruning, the

decoder obtains a maximum speedup at the same

performance level.

The rest of the paper is organized as follows. In

Section 2, we present the dynamic beam threshold-

ing with some statistics observations. In Section 3,

we describe how to calculate the language model

look-ahead probability and incorporate it into the

beam thresholding. In Section 4, we compare our

work to previous work. Section 5 presents the eval-

uation results and comparison curves using different

beam thresholding methods. Finally, we conclude

and discuss our future work in Section 6.

Experimental re-

By combin-

2 Dynamic Beam Thresholding

Generally speaking, if we use a loose beam thresh-

old by retaining as many hypotheses as possible, de-

coding will be very slow although translation qual-

ity remains high. On the other hand, if we use a

tight beam threshold, pruning as many hypotheses

as possible, we can get a considerable speedup but

at the cost of much degraded translation quality. So

the question is how we can find an appropriate beam

threshold to get the best tradeoff between translation

quality and speed. Unfortunately, we are not able

to find such an ideal beam threshold since we don’t

know exactly the search space beforehand.

Most researchers select the beam threshold empir-

ically on the development set and use it constantly

for the whole test set. We call this strategy fixed

beam thresholding (FBT). To guarantee the transla-

tion quality, a loose beam threshold is usually used

at the cost of slow decoding speed. Using such a

constant loose beam threshold on the non-uniformly

distributed search space will waste decoding efforts

for search areas where the decoder collects more ac-

curate information.

A better strategy is to dynamically adjust the

beam threshold based on a hidden variable, which

to some extent associates with the search space dis-

tribution. Here we define the variable as a ratio

r (seq/sent) between the length of source sequence

covered by partial hypotheses and that of the whole

sentence to be translated. To investigate how we

should vary the beam threshold with the length ra-

tio r, we trace the cost1difference (best-corr) be-

tween the best hypothesis and the correct hypothe-

sis2in chart cells on the NIST MT-02 test set (878

sentences, 19.6 words per sentence) which is de-

coded using a very loose beam threshold3. We plot

the curve of average best-corr cost difference vs.

seq/sent length ratio in Figure 1, which visualizes

how wide we should set the beam so that correct hy-

potheses fall inside the beam.

From this figure, we can observe that in most

cases, the longer the source fragment covered by

hypotheses, the smaller the cost difference between

the correct hypotheses and the best hypotheses. This

means that we can safely use a tighter beam thresh-

old for hypotheses covering longer source frag-

ments.It is safe because correct hypotheses are

still included in the beam while incorrect hypotheses

are pruned as many as possible. However, for hy-

potheses covering shorter fragments, we should use

a looser beam threshold to include all possible can-

didates for future exploration so that potential can-

didates can survive to become part of the finally best

hypothesis.

According to this observation, we dynamically

adjust the beam threshold parameter α as a function

1The cost of a hypothesis is the negative logarithm of the

translation probability of it. The higher the probability, the

lower the cost.

2The correct hypothesis is the hypothesis that is part of the

best translation generated by the decoder. The best hypothesis

is the hypothesis with the least cost in the current span. Note

that the best hypothesis is not always the correct hypothesis.

3Here we loosened the beam threshold gradually until the

BLEU score is not changing. Then we used the last beam

threshold we tried.

Page 3

0.0? 0.2? 0.4? 0.6?0.8?1.0?

-0.02?

0.00?

0.02?

0.04?

0.06?

0.08?

0.10?

0.12?

0.14?

0.16?

average best-corr cost difference?

seq/sent length ratio?

Figure 1: Average Best-corr Cost Difference vs. Seq/sent

Length Ratio on the NIST MT-02.

of the length ratio:

α = α0+ (1 − α0) · r

where α0is the initial value of the beam threshold

parameter which is purposely set to a small value

to capture most of the candidates during the early

stage of decoding. We call this pruning strategy dy-

namic beam thresholding (DBT). DBT increases the

parameter α to tighten the beam when more source

words covered. In theory, DBT runs faster than tra-

ditional beam thresholding FBT at the same perfor-

mance level, as our experiments attest.

3Language Model Look-ahead

In traditional beam thresholding used in SMT de-

coding, only the probability estimated from inside

a partial hypothesis is used. This probability does

not give information about the probability of the hy-

pothesis in the context of the complete translation.

In A* decoding for SMT (Och et al., 2001; Zhang

and Gildea, 2006), different heuristic functions are

used to estimate a “future” probability for complet-

ing a partial hypothesis. In CKY bottom-up parsing,

(Goodman, 1997) introduces a prior probability into

the beam thresholding. All of these probabilities are

capable of capturing the outside context interaction,

to some extent.

In this paper, we discuss the LM look-ahead

(LMLA) and examine the question of whether, given

the complicated reordering in SMT, the LM look-

ahead can obtain a considerable speedup in SMT

decoding.The basic idea of the LM look-ahead

is to incorporate the language model interaction of

the boundary words of a hypothesis and neighboring

words outside the hypothesis on the target side into

the pruning process as early as possible. Since the

exact neighboring words are not available until the

partial hypothesis is completed, we obtain potential

neighboring words in two steps as follows.

Firstly, for each sequence of source words si...sj,

we find its most probable translation T(si...sj) with

a monotone search through translation options, only

considering the translation model and the language

model cost. This can be quickly done with dynamic

programming, similar to (Koehn, 2004). Then we

cache the leftmost/rightmost target boundary words

Tl(si...sj)/Tr(si...sj), which both include m?=

min(m − 1,|T(si...sj)|) (m is the language model

order) words4. Since there are only n(n+1)/2 con-

tinuous sequences for n words, the target boundary

words for all these sequences can be quickly found

and cached before decoding with a very cheap over-

head.

Secondly,for a hypothesis H

sourcespan

si...sj, we

most/rightmost target boundary words of its

neighboring spans:

Tl(s1...si−1)/Tr(s1...si−1)

and Tl(sj+1...sn)/Tr(sj+1...sn), which are cached

in the first step. Although these boundary words

are not exactly adjacent to H since there exist

thousands of word reorderings, they still provide

context information for language model interaction.

We utilize them according to the following two

reorderings.

If a straight order is preferred (Fig. 2(a)), the lan-

guage model look-ahead probability πs(H) can be

estimated as follows

covering a

thelook upleft-

πs(H) = m-gram(Tr(s1...si−1),Hl)

·m-gram(Hr,Tl(sj+1...sn))

where Hl/rare the leftmost/rightmost boundary

words of H, which both include m?= min(m −

4The reason for caching m?words is to keep the same with

what we do for each hypothesis, where m?words are also stored

on the left/right of the hypothesis for the dynamic programming

to compute new m-grams in the CKY algorithm intersected

with an m-gram language model (Huang et al., 2005).

Page 4

(b)

source

target

ijj+1

n

1

i-1

(a)

source

target

ij j+1

n

1

i-1

Figure 2: Two Reorderings (straight and inverted) for

Language Model Look-Ahead.

1,|H|) words. If an inverted order is preferred (Fig.

2(b)), the language model look-ahead probability

πi(H) can be estimated as follows

πi(H) = m-gram(Tr(sj+1...sn),Hl)

·m-gram(Hr,Tl(s1...si−1))

Since we don’t know which order will be preferred,

we take the maximum of the straight and inverted

LM look-ahead probability for the hypothesis

π(H) = max(πs(H),πi(H))

The final beam thresholding measure for H when

compared to the best hypothesis within the same cell

is

p(H) = pin(H) · π(H)λLM

where pin(H) is the probability estimated from in-

side the hypothesis H, λLMis the weight of the lan-

guage model. Note that this probability is only used

for the beam thresholding.

4Comparison to Previous Work

Efficient decoding is of great importance to rapid

SMT development and commercial applications.

Much of previous work focuses on reducing the

overwhelming overhead introduced by the in-

tersection of the m-gram language model and

the translation model (phrase-based or syntax-

based).This is the fundamental motivation

for cube pruning/growing(Chiang, 2007; Huang

and Chiang, 2007), and multi-pass decoding ap-

proaches(Venugopal et al., 2007; Zhang and Gildea,

2008). Other efforts have been made for A* decod-

ing using search heuristics (Och et al., 2001; Zhang

and Gildea, 2006).

The Pharaoh decoder (Koehn, 2004) uses an es-

timated score of uncovered source sequences as an

important component to compare hypotheses.

A* decoding (Och et al., 2001; Zhang and Gildea,

2006), a heuristic function is used to estimate the

probability to complete a partial hypothesis.

some extent, both are similar to our LMLA proba-

bility. The biggest difference is that we emphasize

the effect of the language model interaction on the

beam thresholding. We neither use the LMLA prob-

ability as a component of priority functions for A*

decoding(theformer), norusetheestimatedscoreof

the full uncovered source sequence for both thresh-

old pruning and histogram pruning (the latter). The

Pharaoh-style “future cost” can not provide any dis-

criminative information for our pruning since we

compare competing hypotheses within the same cell

(This means that theyhave the same future cost). We

remains the same as the Pharaoh decoder to find the

most probable path through translation options for

source words that are not yet translated. But we go

further to take into account the interaction of current

hypotheses and the most probable path for not yet

translated source sequence.

Moore and Quirk (2007) present two modifica-

tions for beam-search decoding, the Pharaoh de-

coder in particular by improving the future cost es-

timation and early pruning out next-phrase transla-

tions. Their success and the high efficiency of our

beamthresholdingmethods(verifiedbyexperiments

in the next section) show that there is much room for

search space reduction in widely-used beam-search

decoding.

In

To

5Experiments

We carried out a series of experiments to examine

the effect of our beam thresholding techniques by

comparing them with the fixed beam thresholding

as well as the cube pruning, and also by combining

all these pruning approaches step by step. We tested

them on a Chinese-to-English system with a CKY-

style decoder.

Page 5

The system is based on the Bracketing Transduc-

tion Grammars (BTG) (Wu, 1997), which uses the

BTG lexical rules (A → x/y) to translate source

phrase x into target phrase y and the BTG merg-

ing rules (A → [A,A]|?A,A?) to combine two

neighboring phrases with a straight or inverted or-

der. The BTG lexical rules are weighted with several

features, such as phrase translation, word penalty

and language model, in a log-linear form. For the

merging rules, a MaxEnt-based reordering model

usingboundarywordsofneighboringphrasesasfea-

tures is used to predict the merging order, similar

to (Xiong et al., 2006). All the log-linear model

weights are tuned on the development set to maxi-

mize the BLEU score. A CKY-style decoder is de-

veloped to generate the best BTG binary tree for

each input sentence, which yields the best transla-

tion.

We used the FBIS corpus (7.06M Chinese words

and 9.15M English words) as our bilingual training

data, from which a MaxEnt-based reordering model

was also trained. The 4-gram language model train-

ing data (181.1M words) consists of English texts

mostly derived from Xinhua section of the English

Gigaword corpus. We used the NIST MT-05 as our

test set (27.4 words per sentence) and the NIST MT-

02 as our development set.

5.1Dynamic Beam Thresholding

We firstly carried out experiments to compare

the dynamic beam thresholding to the fixed beam

thresholding. By varying the beam threshold (his-

togram pruning parameter b set to 40, beam thresh-

old α for FBT and α0for DBT varying from 0.9

to 0.05), we plot curves of BLEU score vs. decod-

ing time, and BLEU score vs. search space mea-

sured by hypotheses explored per sentence in Figure

3(a) and 3(b) respectively. At most levels of BLEU

score, the speedup is about a factor of 2-3. This im-

provement is highly valuable given that the dynamic

beam thresholding can be implemented without ef-

fort in most SMT systems. The effect of DBT on

the search space reduction is also significant, which

can be observed from Figure 3(b). Usually, 60% of

hypotheses explored by the fixed beam thresholding

can be removed safely by DBT without losing the

translation quality measured by BLEU score .

5.2 Language Model Look-ahead

We examined the language model look-ahead prun-

ing method on both FBT and DBT. The curves of

BLEU score vs. decoding time and BlEU score vs.

search space are plotted in Figure 4(a) and 4(b), un-

der various beam settings in the same fashion of Fig-

ure 3. The “+LMLA” item denotes that the proba-

bility pin· πλLMis used for beam thresholding. In

both cases, the language model look-ahead achieves

significant speedups in terms of decoding time (Fig-

ure 4(a)), a factor of 2-3 for FBT and 1.5-2.5 for

DBT. Figure 4(b) displays larger relative speedups

in terms of search space size, for instance, a fac-

tor of 4-5 for FBT. This is because the language

model look-ahead introduces an additional overhead

for calculating language model probabilities. Al-

though there exists such an overhead for the lan-

guage model look-ahead, it still runs faster than that

without it. The combination of DBT and LMLA

achieves a speedup of a factor 3-5 in terms of de-

coding time when compared to the traditional fixed

beam thresholding.

5.3 Comparison to Cube Pruning

Cube pruning (CP) (Chiang, 2007; Huang and Chi-

ang, 2007) is a state-of-the-art pruning approach

which prunes out a large fraction of the possible hy-

potheses without computing them. We incorporated

this technique in our system and compared it with

our beam thresholding methods.

Figure 5 plots curves of BLEU score vs. decod-

ing time and search space for cube pruning, fixed

beam thresholding, dynamic beam thresholding and

its combination with language model look-ahead.

The curves of cube pruning and our thresholding

method (DBT+LMLA) overlap in most points in

both figures, which displays that the combination

of the dynamic beam thresholding and the language

model look-ahead leads to a speedup improvement

comparable to that of cube pruning.

5.4 Combining with Cube Pruning

The cube pruning algorithm uses beam thresholding

and histogram pruning to examine those survived.

One interesting question is whether we can combine

ourbeamthresholdingmethodsandcubepruningto-

gethertoobtainamaximumspeedup. Wecarriedout

Page 6

0.22

0.23

0.24

0.25

0.26

0.27

0.28

0.1 1

seconds per sentence

10 100 1000

BLEU score

FBT

DBT

(a)

0.22

0.23

0.24

0.25

0.26

0.27

0.28

103

104

105

106

107

BLEU score

hypotheses explored per sentence

FBT

DBT

(b)

Figure 3: Dynamic Beam Thresholding vs. Fixed Beam Thresholding.

0.22

0.23

0.24

0.25

0.26

0.27

0.28

0.1 1

seconds per sentence

10 100 1000

BLEU score

FBT

DBT

FBT+LMLA

DBT+LMLA

(a)

0.22

0.23

0.24

0.25

0.26

0.27

0.28

103

104

105

106

107

BLEU score

hypotheses explored per sentence

FBT

DBT

FBT+LMLA

DBT+LMLA

(b)

Figure 4: Effect of the Language Model Look-Ahead.

0.22

0.23

0.24

0.25

0.26

0.27

0.28

0.1 1

seconds per sentence

10 100 1000

BLEU score

FBT

DBT

CP

DBT+LMLA

(a)

0.22

0.23

0.24

0.25

0.26

0.27

0.28

103

104

105

106

107

BLEU score

hypotheses explored per sentence

FBT

DBT

CP

DBT+LMLA

(b)

Figure 5: Our Beam Thresholding vs. Cube Pruning.

Page 7

0.23

0.24

0.25

0.26

0.27

0.28

0.1 1

seconds per sentence

10 100 1000

BLEU score

CP

CP+LMLA

CP+DBT

(a)

0.23

0.24

0.25

0.26

0.27

0.28

103

104

105

106

107

BLEU score

hypotheses explored per sentence

CP

CP+LMLA

CP+DBT

(b)

Figure 6: Combining with Cube Pruning.

experiments to investigate this combination. Figure

6(a) and 6(b) plot the comparison curves.

We observe that the dynamic beam thresholding

can improve cube pruning, although the speedup is

not as significant as that when compared with the

fixed beam thresholding.

model look-ahead seems not to be helpful for cube

pruning. Figure 6(b) shows that LMLA reduces the

search space of cube pruning only at lower perfor-

mance levels (tight beam settings used) and almost

remains the same size of the search space as that of

cube pruning without LMLA at higher levels. This

indicates that our language model look-ahead is not

sophisticated to provide additional information for

cube pruning since the hypotheses which escape be-

ing pruned by cube pruning also successfully pass

the examination of LMLA.

However, the language

6 Discussion and Future Work

We have presented two efficient beam thresholding

methods to speed up beam-search based SMT de-

coding. We introduce the dynamic beam threshold-

ing based on the observation that most hypotheses

covering a longer source sequence can be pruned

safely with a tighter beam threshold. We use the

language model look-ahead probability to incor-

porate the language model interaction of outside

context with current hypotheses into beam thresh-

olding so that inferior hypotheses can be pruned

early. Both methods lead to significant speed im-

provements when compared to the traditional beam

thresholding. We also combine these two methods

together, which achieves a speed improvement com-

parable to that of cube pruning. The dynamic beam

thresholding is even helpful for the cube pruning.

Although we discuss and use these threshold-

ing methods on BTG-based SMT, they can be ap-

plied to other frameworks, such as standard phrase-

based SMT (Koehn et al., 2003) and linguistically

syntax-based SMT using CKY algorithms. The dy-

namicbeamthresholdingcanbeintegratedintomost

beam-search based SMT decoders without much ef-

forts. However, the language model look-ahead is

much more difficult to be implemented in decod-

ing algorithms using the IBM constraint (Zens et

al., 2004) or dealing with gaps. This remains an

open problem for future research since reorderings

in these decoding algorithms are more complicated.

Although setting the beam threshold as a func-

tion is not a new idea, to our best knowledge, we

are the first to use it in SMT decoding. The func-

tion which we use for the dynamic beam thresh-

olding is actually very straightforward. In the fu-

ture, we will investigate more sophisticated function

for this method. For the look-ahead technique, our

future work includes: 1) calculating more accurate

language model look-ahead probability; 2) investi-

gating other look-ahead techniques beyond LMLA,

such as translation model look-ahead and reorder-

ing model look-ahead and their combination; 3) and

finally applying the look-ahead technique to other

decoding algorithms mentioned above.

Page 8

References

David Chiang. 2007. Hierarchical Phrase-based Transla-

tion. Computational Linguistics, 33(2).

Joshua Goodman. 1997.

Multiple-pass Parsing. In Proceedings of the Sec-

ondConferenceonEmpiricalMethodsinNaturalLan-

guage Processing, pages 11-25.

LiangHuangandDavidChiang. 2007. ForestRescoring:

Faster Decoding with Integrated Language Models. In

Proceedings of the Annual Meeting of the Association

for Computational Linguistics 2007.

Liang Huang, Hao Zhang and Daniel Gildea. 2005. Ma-

chine Translation as Lexicalized Parsing with Hooks.

In Proceedings of the 9th International Workshop

on Parsing Technologies (IWPT-05), Vancouver, BC,

Canada.

Robert C. Moore and Chris Quirk.

Beam-search Decoding for Phrasal Statistical Machine

Translation. In Proceedings of MT-SUMMIT 2007.

Franz Josef Och, Nicola Ueffing, and Herman Ney. 2001.

An Efficient A* Search Algorithm for Statistical Ma-

chine Translation. In Proceedings of ACL Workshop

on Data-Driven Machine Translation.

Kishore Papineni, Salim Roukos, Todd Ward, and Wei-

JingZhu. 2002. BLEU:aMethodforAutomaticEval-

uation of Machine Translation. In Proceedings of the

Annual Meeting of the Association for Computational

Linguistics 2002. Philadelphia, Pennsylvania, USA.

Koehn Philipp, Franz Josef Och, and Daniel Marcu.

2003. Statistical Phrase-Based Translation. In Pro-

ceedings of the Human Language Technology Confer-

ence of the North American Chapter of the Associ-

ation for Computational Linguistics (pp. 127–133).

Edmonton, Alberta, Canada.

Koehn Philipp.2004. PHARAOH: a Beam Search

Decoder for Phrase-Based Statistical Machine Trans-

lation Models, User Manual and Description for

Version 1.2.USC Information Sciences Institute.

http://www.isi.edu/publications/licensedsw/pharaoh/m

anual-v1.2.ps

Ashish Venugopal, Andreas Zollmann and Stephan Vo-

gel.2007.An Efficient Two-Pass Approach to

Synchronous-CFG Driven Statistical MT. In Proceed-

ings of the Human Language Technology and North

American Association for Computational Linguistics

Conference (HLT/NAACL), Rochester, NY. April 22-

27.

Dekai Wu. 1996. A Polynomial-Time Algorithm for

Statistical Machine Translation. In Proceedings of the

Annual Meeting of the Association for Computational

Linguistics 1996.

Dekai Wu.1997.Stochastic Inversion Transduction

Grammars and Bilingual Parsing of Parallel Corpora.

Computational Linguistics, 23(3):377-403.

Global Thresholding and

2007.Faster

Deyi Xiong, Qun Liu and Shouxun Lin. 2006. Maxi-

mum Entropy Based Phrase Reordering Model for Sta-

tistical Machine Translation. In Proceedings of ACL-

COLING 2006.

R. Zens, H. Ney, T. Watanabe, and E. Sumita. 2004. Re-

ordering Constraints for Phrase-Based Statistical Ma-

chine Translation. In Proceedings of COLING 2004,

Geneva, Switzerland, pp. 205-211.

Hao Zhang and Daniel Gildea. 2006. Efficient Search

for Inversion Transduction Grammar. In Proceedings

of ACL-COLING 2006.

Hao Zhang and Daniel Gildea. 2008. Efficient Multi-

pass Decoding for Synchronous Context Free Gram-

mar. In Proceedings of the Annual Meeting of the As-

sociation for Computational Linguistics 2008.