TSD2013.AUTOMATIC MACHINE TRANSLATION EVALUATION WITH PART-OF-SPEECH INFORMATION

Automatic Machine Translation Evaluation with
Part-of-Speech Information
Aaron L.-F. Han, Derek F. Wong, Lidia S. Chao, and Liangye He
University of Macau, Department of Computer and Information Science
Av. Padre Toms Pereira Taipa, Macau, China
{hanlifengaaron}@gmail.com, {derekfw,lidiasc}@umac.mo
{wutianshui0515}@gmail.com
Abstract. One problem of automatic translation is the evaluation of the result.
The result should be as close to a human reference translation as possible, but
varying word order or synonyms have to be taken into account for the evalua-
tion of the similarity of both. In the conventional methods, researchers tend to
employ many resources such as the synonyms vocabulary, paraphrasing, and text
entailment data, etc. To make the evaluation model both accurate and concise, this
paper explores the evaluation only using Part-of-Speech information of the word-
s, which means the method is based only on the consilience of the POS strings of
the hypothesis translation and reference. In this developed method, the POS also
acts as the similar function with the synonyms in addition to its syntactic or mor-
phological behaviour of the lexical item in question. Measures for the similarity
between machine translation and human reference are dependent on the language
pair since the word order or the number of synonyms may vary, for instance. This
new measure solves this problem to a certain extent by introducing weights to
different sources of information. The experiment results on English, German and
French languages correlate on average better with the human reference than some
existing measures, such as BLEU, AMBER and MP4IBM1.1
Keywords: Natural language processing, Machine translation evaluation, Part-
of-Speech, Reference translation
1 Introduction
With the rapid development of Machine Translation systems, 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 commonly used automatic evaluation metrics include the
word error rate WER [2], BLEU [3] (the geometric mean of n-gram precision by the
system output with respect to reference translations), and NIST [4]. Recently, many
other methods were proposed to revise or improve the previous works.
METEOR [5] metric conducts a ﬂexible matching, considering stems, synonyms
and paraphrases, which method and formula for computing a score is much more com-
plicated than BLEU’s [1]. The matching process involves computationally expensive
1
The ﬁnal publication is available at http://www.springerlink.com

2 A.L.F. Han, D.F. Wong, L.S. Chao, L. He
word alignment. There are some parameters such as the relative weight of recall to pre-
cision, the weight for stemming or synonym that should be tuned. Snover [6] discussed
that one disadvantage of the Levenshtein distance was that mismatches in word order
required the deletion and re-insertion of the misplaced words. They proposed TER by
adding an editing step that allows the movement of word sequences from one part of
the output to another. AMBER [7] including AMBER-TI and AMBER-NL declare a
modified version of BLEU and attaches more kinds of penalty coefficients, combining
the n-gram precision and recall with the arithmetic average of F-measure. F15 [8] and
F15G3 perform evaluation with the F1 measure (assigning the same weight on preci-
sion and recall) over target features as a metric for evaluating translation quality. The
target features they defined include TP (be the true positive), TN (the true negative), FP
(the false positive), and FN (the false negative rates), etc. To consider the surrounding
phrases for a missing token in the translation they employed the gapped word sequence
kernels [9] approach to evaluate translations. Other related works include [10], [11] and
[12] about the discussion of word order, ROSE [13], MPF and WMPF [14] about the
employing of POS information, MP4IBM1 [15] without relying on reference transla-
tions, etc.
The evaluation methods proposed previously tend to rely on too many linguistic fea-
tures (difficult in replicability) or no linguistic information (leading the metrics result
in low correlation with human judgments). To address this problem, this paper explores
the performance of a novel method only using the consilience of the POS strings of the
hypothesis translation and reference translation. This ensures that the linguistic infor-
mation is considered in the evaluation but it is a very concise model.
2 Linguistic Features
As discussed above, language variability results in no single correct translation and dif-
ferent languages do not always express the same content in the same way. To address the
variability phenomenon, researchers used to employ the synonyms, paraphrasing or text
entailment as auxiliary information. All of these approaches have their advantages and
weaknesses, e.g. the synonyms are difficult to cover all the acceptable expressions. In-
stead, in the designed metric, we use the part-of-speech (POS) information (also applied
by ROSE [13], MPF and WMPF [14]). If the translation sentence of system outputs is
a good translation then there is a potential that the output sentence has a similar seman-
tic information with the reference sentence (the two sentences may not contain exactly
the same words but with the words that have similar semantic meaning). For example,
“there is a big bag” and “there is a large bag” could be the same expression since “big”
and “large” has the similar meaning (with POS as adjective). To try this approach, we
conduct the evaluation on the POS of the words instead of the words themselves and we
do not use other external resources such as synonym dictionaries. We also test the ap-
proach by calculating the correlation score of this method with human judgments in the
experiment. Assume that we have two sentences: one reference and one system output
translation. Firstly, we extract the POS of each word. Then, we calculate the similarity
of these two sentences through the alignment of their POS information.

Machine Translation Evaluation Based on POS 3
3 Calculation Methods
3.1 Design of hLEPOR Metric
First, we introduce the mathematical harmonic mean for multi-variables (n variables
(X1, X2, ..., Xn)).
Harmonic(X1, X2, ..., Xn) =
n
n
i=1
1
Xi
(1)
where n means the number of variables (also named as factors). Then, the weighted
harmonic mean for multi-variables is:
Harmonic(wX1
X1, wX2
X2, ..., wXn
Xn) =
n
i=1 wXi
n
i=1
wXi
Xi
(2)
where wXi
presents the weight assigned to the corresponding variable Xi. Finally, the
proposed evaluation metric hLEPOR (tunable Harmonic mean of Length Penalty, Pre-
cision, n-gram Position difference Penalty and Recall) is designed as:
hLEPOR = Harmonic(wLP LP, wNP osP enalNPosPenal, wHP RHPR) (3)
=
n
i=1 wi
n
i=1
wi
F actori
=
wLP + wNP osP enal + wHP R
wLP
LP + wNP osP enal
NP osP enal + wHP R
HP R
where LP, NPosPenal and HPR are three factors in hLEPOR and will be in-
troduced in the following. Three tunable weights parameters wLP , wNP osP enal and
wHP R are assigned to the three factors respectively.
3.2 Design of Internal Factors
Length Penalty In the Eq. (3), LP means Length penalty to embrace the penalty for
both longer and shorter system outputs compared with the reference translations:
LP =



e1− r
c : c < r
1 : c = r
e1− c
r : c > r
(4)
where c and r mean the sentence length of candidate translation and reference transla-
tion respectively.
N-gram Position Difference Penalty In the Eq.(3), the NPosPenal is deﬁned as:
NPosPenal = e−NP D
(5)

where NPD means n-gram position difference penalty. The NPosPenal value is de-
signed to compare the POS order in the sentences between reference translation and
output translation. The NPD is defined as:
NPD =
1
Lengthoutput
Lengthoutput
i=1
|PDi| (6)
where Lengthoutput represents the length of system output sentence and PDi means
the n-gram position difference value of aligned POS between output and reference sen-
tences. Every POS from both output translation and reference should be aligned only
once (one-to-one alignment). When there is no match, the value of PDi will be zero as
default for this output POS.
To calculate the NPD value, there are two steps: aligning and calculating. To begin
with, the context-dependent n-gram alignment task: we use the n-gram method and
assign higher priority on it, which means we take into account the surrounding context
(surrounding POS) of the potential POS to select a better matching pairs between the
output and the reference. If there are both nearby matching or there is no matched POS
around the potential pairs, then we consider the nearest matching to align as a backup
choice. The alignment direction is from output sentence to the references.
See example in Figure 1. In the second step (calculating step), we label each POS
with its position number divided by the corresponding sentence length for normaliza-
tion, and then using the Eq. (6) to finish the calculation.
Fig. 1. Example of n-gram POS alignment
We also use the example in Figure 1 for the NPD introduction (Figure 2). In the
example, when we label the position number of output sentence we divide the numerical
position (from 1 to 5) of the current POS by the sentence length 5. For the reference
sentence it is the similar step. After we get the NPD value, using the Eq. (5), the values
of NPosPenal are calculated.
Precision and Recall Precision is designed to reflect the accurate rate of outputs while
recall means the loyalty to the references. In the Eq. (3), HPR means the weighted
Harmonic mean of precision and recall i.e. Harmonic(αR, βP), with parameters α
and β as the tunable weights for recall and precision respectively.
Harmonic(αR, βP) =
α + β
α
R + β
P
(7)

Fig. 2. Example of NPD calculation
P =
alignednum
systemlength
(8)
R =
alignednum
referencelength
(9)
where alignednum represents the number of successfully aligned (matched) POS ap-
pearing both in translation and reference, systemlength and referencelength specify
the sentence length of system output and reference respectively.
System-level hLEPOR We have introduced the calculation of hLEPOR on single
output sentence, and we should consider a proper way to calculate the value when the
cases turn into document (or system) level. We perform the system-level hLEPOR as
below.
hLEPORsys = Harmonic(wLP LPsys, wNP osP enalPosPenaltysys, wHP RHPRsys)
(10)
As shown in the formula, to calculate the system-level score hLEPORsys, we
should firstly calculate the system-level scores of its factors LPsys, PosPenaltysys
and HPRsys. The system level factor scores are calculated by the arithmetic means of
the corresponding sentence-level factor scores.
4 Experiments
We trained hLEPOR and tuned the parameters on the public ACL WMT 20082
data.
There are five languages in the WMT 2008 data including English, Spanish, German,
French and Czech; however, we currently did not find proper parser tools for the Spanish
and Czech languages. So we tested on the English, German and French languages using
the Berkeley parsers [16] to extract the POS information of the tested sentences. Thus,
there are four language pairs in our tested corpora: from German and French to English,
and the inverse. The parameter values on all language pairs are shown in Table 1.
2
http://www.statmt.org/wmt08/

Table 1. Values of tuned parameters
Parameters
(α, β) n-gram POS Alignment Weights(HPR:LP:NPosPenal)
(9,1) 2-gram 3:2:1
The tested corpora we used are from ACL WMT 20113
. There are more than one
hundred MT systems offering their output translation results, with most MT systems
statistical-based except for five rule-based ones. The gold standard reference data for
those corpora consists of 3003 sentences. For each language pairs, there are different
numbers of participated MT systems. Automatic MT evaluation systems are differed by
calculating their Spearman rank correlation coefficient with the human judgment results
[17].
Table 2. Correlation coefficients with human judgments
Correlation Score with Human Judgment
Other-to-English English-to-Other
Metrics DE-EN FR-EN EN-DE EN-FR Mean score
hLEPOR 0.83 0.74 0.84 0.82 0.81
MPF 0.69 0.87 0.63 0.89 0.77
WMPF 0.66 0.87 0.61 0.89 0.76
AMBER-TI 0.63 0.94 0.54 0.84 0.74
AMBER 0.59 0.95 0.53 0.84 0.73
AMBER-NL 0.58 0.94 0.45 0.83 0.7
METEOR-1.3 0.71 0.93 0.3 0.85 0.70
ROSE 0.59 0.86 0.41 0.86 0.68
BLEU 0.48 0.85 0.44 0.86 0.66
F15G3 0.48 0.88 0.3 0.84 0.63
F15 0.45 0.87 0.19 0.85 0.59
MP4IBM1 0.56 0.08 0.91 0.61 0.54
TER 0.33 0.77 0.12 0.84 0.52
We compare the experiments results with several classic metrics including BLEU,
METEOR, TER and some latest ones (e.g. MPF, ROSE, F15, AMBER, MP4IBM1).
The system level correlation coefficients of these metrics with human judgments are
shown in the Table 2 which is ranked by the mean correlation scores of the metrics on
four language pairs. Several conclusions from the results could be drawn: first, many
evaluation metrics performed well in certain language pairs but weak on others, e.g.
WMPF results in 0.89 correlation with human judgments on English-to-French corpus
but down to 0.61 score on English-to-German, F15 gets 0.87 score on French-to-English
but 0.45 on German-to-English, ROSE performs well on both French-to-English and
3
http://www.statmt.org/wmt11/

English-to-French but worse on Germen-to-English and English-to-German. Second,
recently proposed evaluation metrics (e.g. MPF and AMBER) generally perform better
than the traditional ones (e.g. BLEU and TER), showing an improvement of the research
work.
5 Conclusion and Perspectives
To make the evaluation model both accurate and concise, instead of using the syn-
onyms vocabularies, paraphrasing and text entailment that are commonly used by other
researchers, this paper explores the use of the POS information of the words sequences.
What is noticing is that some researchers have used the n-gram method on the words
alignment (e.g. BLEU using 1-gram to 4-gram), other researchers used the POS in-
formation in the similarity calculation by counting the number of corresponding POS.
However, this paper employs the n-gram method on the POS alignment. Since this
developed metric only relies on the consilience of the POS strings of the evaluated sen-
tence even without using the surface words, it has a potential to be further developed
as a reference independent metric. The main difference of this paper and our previous
work LEPOR (the product of factors, perform on words) [18] is that this method group-
s the factors based on mathematical weighted harmonic mean, instead of the simple
product of factors, and this method is performed on the POS instead of words. The
overall weighted harmonic mean allows to tune the model neatly according to different
circumstances, and the POS information can act as part of synonyms in addition to the
syntactic or morphological behaviour of the lexical item.
Even though the designed metric has shown promising performances on the test-
ed language pairs (EN to DE and FR, and the inverse direction). There are several
weaknesses of this metric. Firstly, the POS codes are not language independent, so this
method may not work well on the distant languages e.g. English and Japanese. Second-
ly, the parsing accuracy will effect the performance of the evaluation. Thirdly, to make
the evaluation model concise, this work only uses the POS information without consid-
ering the surface words. To address these weaknesses, in the future work, more language
pairs will be tested, other POS generation tools will be explored, and the combination
of surface words and POS will be employed.
The designed source codes of this paper can be freely downloaded for research
purpose, open source online. The source code of hLEPOR measuring algorithm is
available here “https://github.com/aaronlifenghan/aaron-project-hlepor”. “The ﬁnal pub-
lication is available at http://www.springerlink.com”
Acknowledgments. The authors wish to thank the anonymous reviewers for many
helpful comments. The authors are grateful to the Science and Technology Develop-
ment 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.

References
1. Koehn P.: Statistical Machine Translation, (University of Edinburgh), Cambridge University
Press (2010)
2. Su Keh-Yih, Wu Ming-Wen and Chang Jing-Shin: A New Quantitative Quality Measure for
Machine Translation Systems. In Proceedings of the 14th International Conference on Com-
putational Linguistics, pages 433–439, Nantes, France, July (1992)
3. Papineni, K., Roukos, S., Ward, T. and Zhu, W. J.: BLEU: a method for automatic evaluation
of machine translation. In Proceedings of the ACL 2002, pages 311-318, Philadelphia, PA,
USA (2002)
4. Doddington, G.: Automatic evaluation of machine translation quality using n-gram co-
occurrence statistics. In Proceedings of the second international conference on Human Lan-
guage Technology Research, pages 138-145, San Diego, California, USA (2002)
5. Banerjee, S. and Lavie, A.: Meteor: an automatic metric for MT evaluation with high levels
of correlation with human judgments. In Proceedings of ACL-WMT, pages 65-72, Prague,
Czech Republic (2005)
6. Snover, M., Dorr, B., Schwartz, R., Micciulla, L. and Makhoul, J.: A study of translation edit
rate with targeted human annotation. In Proceedings of the Conference of the Association for
Machine Translation in the Americas, pages 223-231, Boston, USA (2006)
7. Chen, B. and Kuhn, R.: Amber: A modiﬁed bleu, enhanced ranking metric. In Proceedings of
ACL-WMT, pages 71-77, Edinburgh, Scotland, UK (2011)
8. Bicici, E. and Yuret, D.: RegMT system for machine translation, system combination, and
evaluation. In Proceedings ACL-WMT, pages 323-329, Edinburgh, Scotland, UK (2011)
9. Taylor, J. Shawe and N. Cristianini: Kernel Methods for Pattern Analysis. Cambridge Univer-
sity Press 2004.
10. Wong, B. T-M and Kit, C.: Word choice and word position for automatic MT evaluation. In
Workshop: MetricsMATR of the Association for Machine Translation in the Americas, short
paper, 3 pages, Waikiki, Hawai’I, USA (2008)
11. Isozaki, H., Hirao, T., Duh, K., Sudoh, K., and Tsukada, H.: Automatic evaluation of trans-
lation quality for distant language pairs. In Proceedings of the 2010 Conference on EMNLP,
pages 944–952, Cambridge, MA (2010)
12. Talbot, D., Kazawa, H., Ichikawa, H., Katz-Brown, J., Seno, M. and Och, F.: A Lightweight
Evaluation Framework for Machine Translation Reordering. In Proceedings of the Sixth ACL-
WMT, pages 12-21, Edinburgh, Scotland, UK (2011)
13. Song, X. and Cohn, T.: Regression and ranking based optimisation for sentence level MT e-
valuation. In Proceedings of the ACL-WMT, pages 123-129, Edinburgh, Scotland, UK (2011)
14. Popovic, M.: Morphemes and POS tags for n-gram based evaluation metrics. In Proceedings
of ACL-WMT, pages 104-107, Edinburgh, Scotland, UK (2011)
15. Popovic, M., Vilar, D., Avramidis, E. and Burchardt, A.: Evaluation without references: IB-
M1 scores as evaluation metrics. In Proceedings of the ACL-WMT, pages 99-103, Edinburgh,
Scotland, UK (2011)
16. Petrov S., Leon Barrett, Romain Thibaux, and Dan Klein: Learning accurate, compact, and
interpretable tree annotation. Proceedings of the 21st ACL, pages 433–440, Sydney, July
(2006)
17. Callison-Bruch, C., Koehn, P., Monz, C. and Zaidan, O. F.: Findings of the 2011 Workshop
on Statistical Machine Translation. In Proceedings of ACL-WMT, pages 22-64, Edinburgh,
Scotland, UK (2011)
18. Han, Aaron L.-F., Wong, Derek F., and Chao, Lidia S.: LEPOR: A Robust Evaluation Metric
for Machine Translation with Augmented Factors. In Proceedings of the 24th International
Conference on Computational Linguistics (COLING 2012): Posters, pages 441–450, Mumbai,
India (2012)

TSD2013.AUTOMATIC MACHINE TRANSLATION EVALUATION WITH PART-OF-SPEECH INFORMATION

More Related Content

What's hot (20)

Similar to TSD2013.AUTOMATIC MACHINE TRANSLATION EVALUATION WITH PART-OF-SPEECH INFORMATION (20)

More from Lifeng (Aaron) Han (20)

Recently uploaded (20)

TSD2013.AUTOMATIC MACHINE TRANSLATION EVALUATION WITH PART-OF-SPEECH INFORMATION