Language-independent Model for Machine Translation Evaluation with Reinforced Factors

Abstract

The conventional machine translation eval-uation metrics tend to perform well on cer-tain language pairs but weak on other lan-guage pairs. Furthermore, some evalua-tion metrics could only work on certain language pairs not language-independent. Finally, no considering of linguistic in-formation usually leads the metrics re-sult in low correlation with human judg-ments while too many linguistic features or external resources make the metric-s complicated and difficult in replicabil-ity. To address these problems, a nov-el language-independent evaluation metric is proposed in this work with enhanced factors and optional linguistic information (part-of-speech, n-grammar) but not very much. To make the metric perform well on different language pairs, extensive factors are designed to reflect the translation qual-ity and the assigned parameter weights are tunable according to the special character-istics of focused language pairs. Experi-ments show that this novel evaluation met-ric yields better performances compared with several classic evaluation metrics (in-cluding BLEU, TER and METEOR) and two state-of-the-art ones including ROSE and MPF. 1

Full text

Language-independent Model for Machine Translation Evaluation with
Reinforced Factors
Aaron Li-Feng Han, Derek F. Wong, Lidia S. Chao, Liangye He
Yi Lu, Junwen Xing, and Xiaodong Zeng
Natural Language Processing & Portuguese-Chinese Machine Translation Laboratory
Department of Computer and Information Science
University of Macau, Macau S.A.R., China
hanlifengaaron@gmail.com, {derekfw, lidiasc}@umac.mo
{wutianshui0515,takamachi660,nlp2ct.anson,nlp2ct.samuel}@gmail.com
Abstract
The conventional machine translation eval-
uation metrics tend to perform well on cer-
tain language pairs but weak on other lan-
guage pairs. Furthermore, some evalua-
tion metrics could only work on certain
language pairs not language-independent.
Finally, no considering of linguistic in-
formation usually leads the metrics re-
sult in low correlation with human judg-
ments while too many linguistic features
or external resources make the metric-
s complicated and difficult in replicabil-
ity. To address these problems, a nov-
el language-independent evaluation metric
is proposed in this work with enhanced
factors and optional linguistic information
(part-of-speech, n-grammar) but not very
much. To make the metric perform well on
different language pairs, extensive factors
are designed to reflect the translation qual-
ity and the assigned parameter weights are
tunable according to the special character-
istics of focused language pairs. Experi-
ments show that this novel evaluation met-
ric yields better performances compared
with several classic evaluation metrics (in-
cluding BLEU, TER and METEOR) and
two state-of-the-art ones including ROSE
and MPF.
1
1 Introduction
The machine translation (MT) began as early as
in the 1950s (Weaver, 1955) and gained a big
1
The final publication is available at http://www.mt-
archive.info/
progress science the 1990s due to the develop-
ment of computers (storage capacity and compu-
tational power) and the enlarged bilingual corpora
(Marino et al., 2006), e.g. (Och, 2003) present-
ed MERT (Minimum Error Rate Training) for log-
linear statistical machine translation (SMT) mod-
els to achieve better translation quality, (Su et al.,
2009) used the Thematic Role Templates model to
improve the translation and (Xiong et al., 2011)
employed the maximum-entropy model etc. The
statistical MT (Koehn, 2010) became mainly ap-
proaches in MT literature. Due to the wide-spread
development of MT systems, the MT evaluation
becomes more and more important to tell us how
well the MT systems perform and whether they
make some progress. However, the MT evaluation
is difficult because some reasons, e.g. language
variability results in no single correct translation,
the natural languages are highly ambiguous and d-
ifferent languages do not always express the same
content in the same way (Arnold, 2003).
How to evaluate each MT system’s quality and
what should be the criteria have become the new
challenges in front of MT researchers. The earliest
human assessment methods include the intelligi-
bility (measuring how understandable the sentence
is) and fidelity (measuring how much information
the translated sentence retains compared to the o-
riginal) used by the Automatic Language Process-
ing Advisory Committee (ALPAC) around 1966
(Carroll, 1966), and the afterwards proposed ad-
equacy (similar as fidelity), fluency (whether the
sentence is well-formed and fluent) and compre-
hension (improved intelligibility) by Defense Ad-
vanced Research Projects Agency (DARPA) of US
(White et al., 1994). The manual evaluations suf-

fer the main disadvantage that it is time-consuming
and thus too expensive to do frequently.
The early automatic evaluation metrics include
the word error rate WER (Su et al., 1992) (edit dis-
tance between the system output and the closest
reference translation), position independent word
error rate PER (Tillmann et al., 1997) (variant of
WER that disregards word ordering), BLEU (Pap-
ineni et al., 2002) (the geometric mean of n-gram
precision by the system output with respect to
reference translations), NIST (Doddington, 2002)
(adding the information weight) and GTM (Turian
et al., 2003). Recently, many other methods were
proposed to revise or improve the previous works.
One of the categories is the lexical similarity
based metric. The metrics of this kind include
the edit distance based method, such as the TER
(Snover et al., 2006) and the work of (Akiba et
al., 2001) in addition to WER and PER, the pre-
cision based method such as SIA (Liu and Gildea,
2006) in addition to BLEU and NIST, recall based
method such as ROUGE (Lin and Hovy, 2003),
the word order information utilized by (Wong and
Kit, 2008), (Isozaki et al., 2010) and (Talbot et
al., 2011), and the combination of precision and
recall such as Meteor-1.3 (Denkowski and Lavie,
2011) (an modified version of Meteor, includes
ranking and adequacy versions and has overcome
some weaknesses of previous version such as noise
in the paraphrase matching, lack of punctuation
handling and discrimination between word type-
s), BLANC (Lita et al., 2005), LEPOR (Han et
al., 2012) and PORT (Chen et al., 2012). An-
other category is the employing of linguistic fea-
tures. The metrics of this kind include the syntactic
similarity such as the Part-of-Speech information
used by ROSE (Song and Cohn, 2011) and MPF
(Popovic, 2011), and phrase information employed
by (Echizen-ya and Araki, 2010) and (Han et al.,
2013b); the semantic similarity such as Textual en-
tailment used by (Mirkin et al., 2009), Synonyms
by (Chan and Ng, 2008), paraphrase by (Snover et
al., 2009).
The evaluation methods proposed previously
suffer from several main weaknesses more or less:
perform well in certain language pairs but weak on
others, which we call the language-bias problem;
consider no linguistic information (not reasonable
from the aspect of linguistic analysis) or too many
linguistic features (making it difficult in replicabil-
ity), which we call the extremism problem; present
incomprehensive factors (e.g. BLEU focus on pre-
cision only). To address these problems, a novel
automatic evaluation metric is proposed in this pa-
per with enhanced factors, tunable parameters and
optional linguistic information (part-of-speech, n-
gram).
2 Designed Model
2.1 Employed Internal Factors
Firstly, we introduce the internal factors utilized in
the calculation model.
2.1.1 Enhanced Length Penalty
Enhanced length penalty ELP is designed to
put the penalty on both longer and shorter sys-
tem output translations (an enhanced version of the
brevity penalty in BLEU):
ELP =
(
e
1−
r
c
: c < r
e
1−
c
r
: c ≥ r
(1)
where the parameters c and r are the sentence
length of automatically output (candidate) and ref-
erence translation respectively.
2.1.2 N-gram Position Difference Penalty
The N-gram Position Difference Penalty
NP osP enal is developed to compare the word
order between the output and reference translation.
NP osP enal = e
−NP D
(2)
where NP D is defined as:
NP D =
1
Length
output
Length
output
X
i=1
|P D
i
| (3)
where Length
output
is the length of system output
sentence and P D
i
means the position difference
value of each output word. Every word from both
output translation and reference should be aligned
only once. When there is no match, the value of
P D
i
is assigned with zero as default for this output
token.
Two steps are designed to measure the NP D
value. The first step is the context-dependent n-
gram alignment: we use the n-gram method and
assign it with higher priority, which means the sur-
rounding context of the potential words are con-
sidered when selecting the matched pairs between

the output and reference sentence. The nearest
match is accepted as a backup choice to establish
the alignment, if there are both nearby matching or
there is no other matched words surrounding the
potential word pairs. The one-direction alignment
is from output sentence to the reference.
Assuming that w
x
represents the current word
in output sentence and w
x+k
means the kth word
to the previous (k < 0) or following (k > 0).
On the other hand, w
r
y
means the word matching
w
x
in the references, and w
r
y+j
has the similar
meaning as w
x+k
but in reference sentence. The
variable Distance is the position difference val-
ue between the matching word in outputs and ref-
erences. The operation process and pseudo code
of the context-dependent n-gram word alignmen-
t algorithm are shown in Figure 1 (with → as the
alignment). There is an example in Figure 2. In
the calculating step, each word is labeled with the
quotient value of its position number divided by
sentence length (the total number of the tokens in
the sentence).
Let’s see the example in Figure 2 for the NP D
introduction (Figure 3). Each output word is la-
beled with the position quotient value from 1/6 to
6/6 (indicating the word position marked by sen-
tence length which is 6). The words in the refer-
ence sentence is labeled using the same subscripts.
2.1.3 Precision and Recall
Precision and recall are two commonly used cri-
teria in the NLP literature. We use the HP R to
represent the weighted Harmonic mean of preci-
sion and recall, i.e. Harmonic(αR, βP ). The
weights are the tunable parameters α and β.
HP R =
(α + β)P recision × Recall
αP recision + βRecall
(4)
P recision =
Aligned
num
Length
output
(5)
Recall =
Aligned
num
Length
ref erence
(6)
where Aligned
num
represents the number of suc-
cessfully matched words appearing both in trans-
lation and reference.
2.2 Sentence Level Score
Secondly, we introduce the mathematical harmon-
ic mean to group multi-variables (n variables
(X
1
, X
2
, . . . , X
n
)).
Harmonic(X
1
, X
2
, ..., X
n
) =
n
P
n
i=1
1
X
i
(7)
where n is the number of factors. Then, the
weighted harmonic mean for multi-variables is:
Harmonic(w
X
1
X
1
, w
X
2
X
2
, ..., w
X
n
X
n
) =
P
n
i=1
w
X
i
P
n
i=1
w
X
i
X
i
(8)
where w
X
i
is the weight of variable X
i
. Final-
ly, the sentence level score of the developed eval-
uation metric hLEP OR (Harmonic mean of en-
hanced Length Penalty, Precision, n-gram Position
difference Penalty and Recall) is measured by:
hLEP OR =
=
P
n
i=1
w
i
P
n
i=1
w
i
F actor
i
=
w
ELP
+ w
NP osP enal
+ w
HP R
w
ELP
ELP
+
w
NP osP enal
NP osP enal
+
w
HP R
HP R
(9)
where ELP , NP osP enal and HP R are the three
factors explained in previous section with tunable
weights w
ELP
, w
NP osP enal
and w
HP R
respective-
ly.
2.3 System-level Score
The system level score is the arithmetical mean of
the sentence scores as below.
hLEP OR =
1
SentNum
SentN um
X
i=1
hLEP OR
i
(10)
where hLEP OR represents the system-level s-
core of hLEP OR, SentNum specifies the
sentence number of the test document, and
hLEP OR
i
means the score of the ith sentence.
3 Enhanced Version
This section introduces an enhanced version of the
developed metric hLEP OR as hLEP OR
E
. As
discussed by many researchers, language variabili-
ty results in no single correct translation and differ-
ent languages do not always express the same con-
tent in the same way. In addition to the augment-
ed factors of the designed metric hLEP OR, we
present that optional linguistic information can be

combined into this metric concisely. As an exam-
ple, we will show how the part-of-speech (POS) in-
formation can be employed into this metric. First,
we calculate the system-level hLEP OR scores
on the surface words ( hLEP OR
word
). Then
we employ the same algorithms of hLEP OR on
the corresponding POS sequences of the words (
hLEP OR
P OS
). Finally, we combine this two
system-level scores together with tunable weights
(w
hw
and w
hp
) as the final score.
hLEP OR
E
=
1
w
hw
+ w
hp
(w
hw
hLEP OR
word
+w
hp
hLEP OR
P OS
) (11)
We mention the POS information because it
sometimes acts as the similar function with the
synonyms, e.g. “there is a big bag” and “there
is a large bag” could be the same meaning but
with different surface words “big” and “large”
(the same POS adjective). The POS information
has been proved helpful in the research works of
ROSE (Song and Cohn, 2011) and MPF (Popovic,
2011). The POS information could be replaced by
any other concise linguistic information in our de-
signed model.
4 Evaluating the Evaluation Metric
In order to distinguish the reliability of differen-
t MT evaluation metrics, Spearman rank correla-
tion coefficient ρ is commonly used to calculate
the correlation in the annual workshop of statisti-
cal machine translation (WMT) for Association of
Computational Linguistics (ACL) (Callison-Burch
et al., 2011). When there are no ties, Spearman
rank correlation coefficient is calculated as:
ρ
φ(XY )
= 1 −
6
P
n
i=1
d
2
i
n(n
2
− 1)
(12)
where d
i
is the difference-value (D-value) be-
tween the two corresponding rank variables
~
X =
{x
1
, x
2
, ..., x
n
} and
~
Y = {y
1
, y
2
, ..., y
n
} describ-
ing the system φ and n is the number of variables
in the system.
5 Experiments
The experiment corpora are from the ACL’s spe-
cial interest group of machine translation SIGMT
(WMT workshop) which contain eight corpora in-
cluding English-to-other (Spanish, Czech, French
and German) and other-to-English. There are in-
deed a lot of linguistic POS tagger tools for differ-
ent languages available. We conduct an evaluation
with different POS taggers, and find that the em-
ploying of POS information can make an increase
of the correlation score with human judgment for
some language pairs but little or no effect on other-
s. The employed POS tagging tools include Berke-
ley POS tagger for French, English and German
(Petrov et al., 2006), COMPOST Czech morphol-
ogy tagger (Collins, 2002) and TreeTagger Span-
ish tagger (Schmid, 1994). To avoid the overfitting
problem, the WMT 2008
2
data are used in the de-
velopment stage for the tuning of the parameter-
s and the WMT 2011 corpora are used in testing.
The tuned parameter values for different language
pairs are shown in Table 1. The abbreviations EN,
CZ, DE, ES and FR mean English, Czech, Ger-
man, Spanish and French respectively. In the n-
gram word (POS) alignment, bigram is selected in
all the language pairs. To make the model con-
cise using as fewer of external resources as possi-
ble, the value of “N/A” means the POS informa-
tion of that language pair is not employed due to
that it makes little or no effect in the correlation
scores. The label “(W)” and “(POS)” means the
parameters tuned on word and POS respectively.
The “NPP” means N P osP enal to save window
space. The tuned parameter values also prove that
different language pairs embrace different charac-
teristics.
The testing results on WMT 2011
3
corpora are
shown in Table 2. The comparisons with language-
independent evaluation metrics include the clas-
sic metrics (BLEU, TER and METEOR) and two
state-of-the-art metrics MPF and ROSE. We selec-
t MPF and ROSE because that these two metrics
also employ the POS information and MPF yield-
ed the highest correlation score with human judg-
ments among all the language-independent metric-
s (performing on eight language pairs) in WMT
2011. The numbers of participated automatic MT
systems in WMT 2011 are 10, 22, 15 and 17 re-
spectively for English-to-other (CZ, DE, ES and
FR) and 8, 20, 15 and 18 respectively for the op-
posite translation direction. The gold standard ref-
erence data for those corpora consists of 3,003 sen-
tences offered by manual work. Automatic MT e-
2
http://www.statmt.org/wmt08/
3
http://www.statmt.org/wmt11/

Ratio Other-to-English English-to-Other
CZ-EN DE-EN ES-EN FR-EN EN-CZ EN-DE EN-ES EN-FR
HPR:ELP:NPP(W) 7:2:1 3:2:1 7:2:1 3:2:1 3:2:1 1:3:7 3:2:1 3:2:1
HPR:ELP:NPP(POS) N/A 3:2:1 N/A 3:2:1 N/A 7:2:1 N/A 3:2:1
α : β(W) 1:9 9:1 1:9 9:1 9:1 9:1 9:1 9:1
α : β(POS) N/A 9:1 N/A 9:1 N/A 9:1 N/A 9:1
w
hw
: w
hp
N/A 1:9 N/A 9:1 N/A 1:9 N/A 9:1
Table 1: Values of tuned weight parameters
valuation metrics are evaluated by the correlation
coefficient with the human judgments.
Several conclusions could be drawn from the re-
sults. First, some evaluation metrics show good
performances on part of the language pairs but low
performances on others, e.g ROSE results in 0.92
correlation with human judgments on Spanish-to-
English corpus but down to 0.41 score on English-
to-German; METEOR gets 0.93 score on French-
to-English but 0.3 on English-to-German. Second,
hLEP OR
E
generally yields good performances
on different language pairs except for the English-
to-Czech and results in the highest Mean correla-
tion score 0.83 on eight corpora. Third, the recent-
ly developed methods (e.g. MPF, 0.81 mean score)
correlate better with human judgments than the tra-
ditional ones (e.g. BLEU, 0.74 means score), indi-
cating an improvement of the researches. Final-
ly, no metric can yield high performance on all
the language pairs, which shows that there remains
large potential to achieve improvement.
6 Conclusion and Perspectives
This work proposes a language-independent mod-
el for machine translation evaluation. Considering
the different characteristics of different languages,
hLEP OR
E
has been extensively designed from
different aspects. That spans from word order
(context-dependent n-gram alignment), output ac-
curacy (precision), and loyalty (recall) to trans-
lation length performance (sentence length). D-
ifferent weight parameters are assigned to adjust
the importance of each factor, for instance, the
word position could be free in some languages but
strictly constrained in other languages. In prac-
tice, these employed features by hLEP OR
E
are
also the vital ones when people facilitate language
translation. This is the philology behind the formu-
lation and the study of this work, and we believe
human’s translation ideology is the exact direction
that MT systems should try to approach. Further-
more, this work specifies that different external re-
sources or linguistic information could be integrat-
ed into this model easily. As suggested by other
works, e.g. (Avramidis et al., 2011), the POS infor-
mation is considered in the experiments and shows
some improvements on certain language pairs.
There are several main contributions of this pa-
per compared with our previous work (Han et al.,
2013). This work combines the utilizing of sur-
face words and linguistic features together (in-
stead of relying on the consilience of the POS se-
quence only). This paper measures the system-
level hLEP OR score by the arithmetical mean of
each sentence-level score (instead of the Harmon-
ic mean of system-level internal factors). This pa-
per shows the performances of enhanced method
hLEP OR
E
on all the eight language pairs re-
leased by WMT official web (instead of part lan-
guage pairs by previous work) and most of the per-
formances have achieved improvements than pre-
vious work on the same language pairs (e.g. the
correlation score on German-English is 0.86 in-
creased from 0.83; the correlation score on French-
English is 0.92 increased from 0.74.). Other po-
tential linguistic features are easily to be employed
into the flexible model built in this paper.
There are also several aspects that should be ad-
dressed in the future works. Firstly, more language
pairs, in addition to the European languages, will
be tested such as Japanese, Korean and Chinese
and the performances of linguistic features (e.g.
POS tagging) will also be explored on the new lan-
guage pairs. Secondly, the tuning of weight pa-
rameters to achieve high correlation with human
judgments during the development period will be
automatically performed. Thirdly, since the use
of multiple references helps the usual translation

Metrics Other-to-English English-to-Other Mean
CZ-EN DE-EN ES-EN FR-EN EN-CZ EN-DE EN-ES EN-FR
hLEP OR
E
0.93 0.86 0.88 0.92 0.56 0.82 0.85 0.83 0.83
MPF 0.95 0.69 0.83 0.87 0.72 0.63 0.87 0.89 0.81
ROSE 0.88 0.59 0.92 0.86 0.65 0.41 0.9 0.86 0.76
METEOR 0.93 0.71 0.91 0.93 0.65 0.3 0.74 0.85 0.75
BLEU 0.88 0.48 0.9 0.85 0.65 0.44 0.87 0.86 0.74
TER 0.83 0.33 0.89 0.77 0.5 0.12 0.81 0.84 0.64
Table 2: Correlation coefficients with human judgments
quality measures correlate with the human judg-
ing, the scheme of how to use the multiple refer-
ences will be designed.
The designed source codes of this paper
can be freely downloaded for research pur-
pose, open source online. The source code of
hLEPOR measuring algorithm is available here
“https://github.com/aaronlifenghan/aaron-project-
hlepor”. “The final publication is available at
http://www.mt-archive.info/”
Acknowledgments.
The authors are grateful to the Science and
Technology Development Fund of Macau and the
Research Committee of the University of Macau
for the funding support for our research, under
the reference No. 017/2009/A and RG060/09-
10S/CS/FST. The authors also wish to thank the
anonymous reviewers for many helpful comments.
References
Akiba, Y., K. Imamura, and E. Sumita. 2001. Using
Multiple Edit Distances to Automatically Rank Ma-
chine Translation Output. Proceedings of MT Sum-
mit VIII , Santiago de Compostela, Spain.
Arnold, D. 2003. Why translation is difficult for com-
puters. In Computers and Translation: A transla-
tor’s guide , Benjamins Translation Library.
Avramidis, E., Popovic, M., Vilar, D., Burchardt, A.
2011. Evaluate with Confidence Estimation: Ma-
chine ranking of translation outputs using grammat-
ical features. Proceedings of ACL-WMT , pages 65-
70, Edinburgh, Scotland, UK.
Callison-Bruch, C., Koehn, P., Monz, C. and Zaidan, O.
F. 2011. Findings of the 2011 Workshop on Statisti-
cal Machine Translation. Proceedings of ACL-WMT
, pages 22-64, Edinburgh, Scotland, UK.
Carroll, J. B. 1966. Aan experiment in evaluating
the quality of translation. Languages and machines:
computers in translation and linguistics , Automat-
ic Language Processing Advisory Committee (AL-
PAC), Publication 1416, Division of Behavioral Sci-
ences, National Academy of Sciences, National Re-
search Council, page 67-75.
Chan, Y. S. and Ng, H. T. 2008. MAXSIM: A maxi-
mum similarity metric for machine translation eval-
uation. Proceedings of ACL 2008: HLT , pages
55–62.
Chen, Boxing, Roland Kuhn and Samuel Larkin.
2012. PORT: a Precision-Order-Recall MT Evalu-
ation Metric for Tuning. Proceedings of 50th ACL) ,
pages 930–939, Jeju, Republic of Korea.
Collins, M. 2002. Discriminative Training Method-
s for Hidden Markov Models: Theory and Experi-
ments with Perceptron Algorithms. Proceedings of
the ACL-02 conference, Volume 10 (EMNLP 02) ,
pages 1-8. Stroudsburg, PA, USA .
Denkowski, M. and Lavie, A. 2011. Meteor: Meteor
1.3: Automatic metric for reliable optimization and
evaluation of machine translation systems. Proceed-
ings of (ACL-WMT) ,pages 85-91, Edinburgh, Scot-
land, UK.
Doddington, G. 2002. Automatic evaluation of ma-
chine translation quality using n-gram co-occurrence
statistics. Proceedings of the second internation-
al conference on Human Language Technology Re-
search , pages 138-145, San Diego, California, USA.
Echizen-ya, H. and Araki, K. 2010. Automatic eval-
uation method for machine translation using noun-
phrase chunking. Proceedings of ACL 2010 , pages
108–117. Association for Computational Linguistic-
s.
Han, Aaron L.-F., Derek F. Wong, and Lidia S. Chao.
2012. LEPOR: A Robust Evaluation Metric for Ma-
chine Translation with Augmented Factors. Pro-
ceedings of the 24th International Conference of
COLING, Posters, pages 441-450, Mumbai, India.
Han, Aaron L.-F., Derek F. Wong, Lidia S. Chao, and
Liangye He. 2013. Automatic Machine Trans-
lation Evaluation with Part-of-Speech Information.

Proceedings of the 16th International Conference of
Text, Speech and Dialogue (TSD 2013), LNCS Vol-
ume Editors: Vaclav Matousek et al. Springer-Verlag
Berlin Heidelberg. Plzen, Czech Republic.
Han, Aaron L.-F., Derek F. Wong, Lidia S. Chao,
Liangye He, Shuo Li, and Ling Zhu. 2013b.
Phrase Mapping for French and English Treebank
and the Application in Machine Translation Evalu-
ation. Proceedings of the International Conference
of the German Society for Computational Linguis-
tics and Language Technology, (GSCL 2013), LNC-
S Volume Editors: Iryna Gurevych, Chris Biemann
and Torsten Zesch. Darmstadt, Germany.
Isozaki, H., Hirao, T., Duh, K., Sudoh, K., and Tsuka-
da, H. 2010. Automatic evaluation of translation
quality for distant language pairs. Proceedings of
the 2010 Conference on EMNLP , pages 944–952,
Cambridge, MA.
Koehn, P. 2010. Statistical Machine Translation. Cam-
bridge University Press .
Marino, B. Jose, Rafael E. Banchs, Josep M. Crego,
Adria de Gispert, Patrik Lambert, Jose A. Fonollosa,
and Marta R. Costa-jussa. 2006. N -gram based ma-
chine translation. Journal of the Computational Lin-
guistics ,Vol. 32, No. 4. pp. 527-549, MIT Press.
Lin, Chin-Yew and E.H. Hovy. 2003. Automatic Eval-
uation of Summaries Using N-gram Co-occurrence
Statistics. Proceedings of HLT-NAACL 2003, Ed-
monton, Canada.
Lita, Lucian Vlad, Monica Rogati and Alon Lavie.
2005. BLANC: Learning Evaluation Metrics for
MT. Proceedings of the HLT/EMNLP, pages
740–747, Vancouver.
Liu D. and Daniel Gildea. 2006. Stochastic iterative
alignment for machine translation evaluation. Pro-
ceedings of ACL-06, Sydney.
Mirkin S., Lucia Specia, Nicola Cancedda, Ido Dagan,
Marc Dymetman, and Idan Szpektor. 2009. Source-
Language Entailment Modeling for Translating Un-
known Terms. Proceedings of the ACL-IJCNLP
2009) , pages 791–799, Suntec, Singapore.
Och, F. J. 2003. Minimum Error Rate Training for S-
tatistical Machine Translation. Proceedings of ACL-
2003 , pp. 160-167.
Papineni, K., Roukos, S., Ward, T. and Zhu, W. J. 2002.
BLEU: a method for automatic evaluation of ma-
chine translation. Proceedings of the ACL 2002 ,
pages 311-318, Philadelphia, PA, USA.
Petrov, S., Leon Barrett, Romain Thibaux, and Dan K-
lein 2006. Learning accurate, compact, and inter-
pretable tree annotation. Proceedings of the 21st A-
CL , pages 433–440, Sydney.
Popovic, M. 2011. Morphemes and POS tags for n-
gram based evaluation metrics. Proceedings of WMT
, pages 104-107, Edinburgh, Scotland, UK.
Schmid, H. 1994. Probabilistic Part-of-Speech Tag-
ging Using Decision Trees. Proceedings of Inter-
national Conference on New Methods in Language
Processing , Manchester, UK.
Snover, M., Dorr, B., Schwartz, R., Micciulla, L. and
Makhoul J. 2006. A study of translation edit rate
with targeted human annotation. Proceedings of the
AMTA, pages 223-231, Boston, USA.
Snover, Matthew G., Nitin Madnani, Bonnie Dorr, and
Richard Schwartz. 2009. TER-Plus: paraphrase, se-
mantic, and alignment enhancements to Translation
Edit Rate. J. Machine Translation, 23: 117-127.
Song, X. and Cohn, T. 2011. Regression and rank-
ing based optimisation for sentence level MT eval-
uation. Proceedings of the WMT , pages 123-129,
Edinburgh, Scotland, UK.
Su, Hung-Yu and Chung-Hsien Wu. 2009. Improving
Structural Statistical Machine Translation for Sign
Language With Small Corpus Using Thematic Role
Templates as Translation Memory. IEEE TRANS-
ACTIONS ON AUDIO, SPEECH, AND LANGUAGE
PROCESSING , VOL. 17, NO. 7.
Su, Keh-Yih, Wu Ming-Wen and Chang Jing-Shin.
1992. A New Quantitative Quality Measure for Ma-
chine Translation Systems. Proceedings of COL-
ING, pages 433–439, Nantes, France.
Talbot, D., Kazawa, H., Ichikawa, H., Katz-Brown, J.,
Seno, M. and Och, F. 2011. A Lightweight Evalu-
ation Framework for Machine Translation Reorder-
ing. Proceedings of the WMT, pages 12-21, Edin-
burgh, Scotland, UK.
Tillmann, C., Stephan Vogel, Hermann Ney, Arkaitz
Zubiaga, and Hassan Sawaf. 1997. Accelerated
DP Based Search For Statistical Translation. Pro-
ceedings of the 5th European Conference on Speech
Communication and Technology .
Turian, J. P., Shen, L. and Melanmed, I. D. 2003. E-
valuation of machine translation and its evaluation.
Proceedings of MT Summit IX , pages 386-393, New
Orleans, LA, USA.
Weaver, Warren. 1955. Translation. Machine Trans-
lation of Languages: Fourteen Essays, In William
Locke and A. Donald Booth, editors, John Wiley and
Sons. New York, pages 15—23.
White, J. S., O’Connell, T. A., and O’Mara, F. E. 1994.
The ARPA MT evaluation methodologies: Evolu-
tion, lessons, and future approaches. Proceedings
of AMTA, pp193-205.

Wong, B. T-M and Kit, C. 2008. Word choice and
word position for automatic MT evaluation. Work-
shop: MetricsMATR of AMTA, short paper, 3 pages,
Waikiki, Hawai’I, USA.
Xiong, D., M. Zhang, H. Li. 2011. A Maximum-
Entropy Segmentation Model for Statistical Machine
Translation. IEEE Transactions on Audio, Speech,
and Language Processing , Volume: 19, Issue: 8,
2011 , pp. 2494- 2505.

Figure 1: N-gram word alignment algorithm
Figure 2: Example of n-gram word alignment
Figure 3: Example of NP D calculation