GSCL2013.Phrase Tagset Mapping for French and English Treebanks and Its Application in Machine Translation Evaluation

Phrase Tagset Mapping for French and English
Treebanks and Its Application in Machine
Translation Evaluation
Aaron Li-Feng Han, Derek F. Wong, Lidia S. Chao
Liangye He, Shuo Li, and Ling Zhu
University of Macau, Department of Computer and Information Science
Av. Padre Toms Pereira Taipa, Macau, China
hanlifengaaron@gmail.com, {derekfw,lidiasc}@umac.mo
{wutianshui0515,leevis1987,imecholing}@gmail.com
Abstract. Many treebanks have been developed in recent years for dif-
ferent languages. But these treebanks usually employ different syntactic
tag sets. This forms an obstacle for other researchers to take full advan-
tages of them, especially when they undertake the multilingual research.
To address this problem and to facilitate future research in unsupervised
induction of syntactic structures, some researchers have developed a u-
niversal POS tag set. However, the disaccord problem of the phrase tag
sets remains unsolved. Trying to bridge the phrase level tag sets of mul-
tilingual treebanks, this paper designs a phrase mapping between the
French Treebank and the English Penn Treebank. Furthermore, one of
the potential applications of this mapping work is explored in the ma-
chine translation evaluation task. This novel evaluation model developed
without using reference translations yields promising results as compared
to the state-of-the-art evaluation metrics. 1
Keywords: Natural language processing, Phrase tagset mapping, Mul-
tilingual treebanks, Machine translation evaluation.
1 Introduction
To promote the development of syntactic analysis technology, treebanks for dif-
ferent languages were developed during the past years, such as the English Penn
Treebank [1][2], the German Negra Treebank [3], the French Treebank [4], the
Chinese Sinica Treebank [5], etc. These treebanks use their own syntactic tagset-
s, with the number of tags ranging from tens (e.g. the English Penn Treebank)
to hundreds (e.g. the Chinese Sinica Treebank). These differences make it incon-
venient for other researchers to take full advantages of the treebanks, especially
when they undertake the multilingual or cross-lingual research. To bridge the
gap between these treebanks and to facilitate future research, such as the unsu-
pervised induction of syntactic structures, Petrov et al. [6] developed a universal
1
The final publication is available at http://www.springerlink.com

2 A.L.F. Han, D.F. Wong, L.S. Chao, L. He, S. Li, L. Zhu
part-of-speech (POS) tagset and the POS tags mapping between multilingual
treebanks. However, the disaccord problem in the phrase level tags remains un-
solved. Trying to bridge the phrase level tags of multilingual treebanks, this paper
designs a mapping work of phrase tags between French Treebank and English
Peen Treeebank (I and II). Furthermore, this mapping work has been applied
into the machine translation evaluation task (one of the potential applications)
to show its advantages.
2 Designed Phrase Mapping
The designed phrase tags mapping between English and French treebanks is
shown in Table 1. There are 9 universal phrasal categories covering the 14 phrase
tags of Peen Treebank I, 26 phrase tags in Penn Treebank II and 14 phrase tags
in the French Treebank. The brief analysis of the phrase tags mapping will be
introduced as below.
Table 1. Phrase tagset mapping for French and English Treebanks
Universal tags English Penn
Treebank I [1]
English Penn
Treebank II [7]
French Tree-
bank [4]
Tags Meaning
NP NP, WHNP NP, NAC, NX,
WHNP, QP
NP Noun phrase
VP VP VP VN, VP, VPpart,
VPinf
Verbal phrase
AJP ADJP ADJP, WHAD-
JP
AP Adjectival phrase
AVP ADVP, WHAD-
VP
ADVP, WHAVP,
PRT
AdP Adverbial phrase
PP PP, WHPP PP, WHPP PP Prepositional
phrase
S S, SBAR, SBAR-
Q, SINV, SQ
S, SBAR, SBAR-
Q, SINV, SQ,
PRN, FRAG,
RRC
SENT, Ssub, Sin-
t, Srel, S
(sub-)Sentence
CONJP CONJP Conjunction
phrase
COP UCP COORD Coordinated
phrase
X X X, INTJ, LST Others
The universal phrase tag NP (noun phrases) covers: the French tag NP (noun
phrases); the English tags NP (noun phrases), NAC (not a constituent, used to
show the scope of certain prenominal modiﬁers within an NP), NX (used within
certain complex NPs to mark the head of the NP), WHNP (wh-noun phrase,

Phrase Tagset Mapping for French and English Treebanks 3
containing some wh-word, e.g. who, none of which) and QP (quantifier phrase).
There are no corresponding French tags that could be aligned to English tag
QP. On the other hand the corresponding quantifier phrase is usually marked
as NP in French, so the English tag QP is classified into the NP category in the
mapping, for instance:
English: (QP (CD 200) (NNS millions))
French: (NP (D 200) (N millions))
The universal phrase tag VP (verb phrase) covers: the French tags VN (verbal
nucleus), VP (infinitives and nonfinite clauses, including VPpart and VPinf); the
English tag VP (verb phrase). In English, the infinitive phrase (to + verb) is
labeled as VP, so the French VP (infinitives and nonfinite clauses) is classified
into the same category, for instance (English infinitive):
(VP (TO to)
(VP (VB come)))
The universal phrase tag AJP (adjective phrase) covers: the French tag AP
(adjectival phrases); the English tags ADJP (adjective phrase), WHADJP (wh-
adjective phrase, e.g. how hot).
The universal phrase tag AVP (adverbial phrase) contains: the French tag
AdP (adverbial phrases); the English tags ADVP (adverb phrase), WHAVP (wh-
adverb phrase) and PRT (particle, a category for words that should be tagged
RP). The English phrase tag PRT is classified into the adverbial phrase tag
because it is usually used to label the word behind an English verb and there is
no similar phrase tag in the French Treebank to align it, for instance (English
PRT):
(VP (VBG rolling)
(PRT (RP up)))
The universal phrase tag PP (prepositional phrase) covers: the French tag PP
(prepositional phrases); the English tags PP (prepositional phrase) and WHPP
(wh-prepositional phrase).
The universal phrase tag S (sentence and sub-sentence) covers: the French
tags SENT (sentences), S (finite clauses, including Ssub, Sint and Srel); the
English tags S (simple declarative clause), SBAR (clause introduced by a sub-
ordinating conjunction), SBARQ (direct question introduced by a wh-word or
a wh-phrase), SINV (declarative sentence with subject-aux inversion), SQ (sub-
constituent of SBARQ excluding wh-phrase), PRN (parenthetical), FRAG (frag-
ment) and RRC (reduced relative clause). The English tag FRAG marks those
portions of text that appear to be clauses but lack too many essential elements
for the exact structure to be easily determined; the RRC label is used only in
cases where “there is no VP and an extra level is needed for proper attachment
of sentential modifiers” [7]. FRAG and RRC tags are at the same level as SBAR,
SINV, and SQ, so they are classified uniformly into S category in the mapping,
for instance:
(FRAG (, ,)
(NP (JJ for) (NN example)))
(RRC

(ADVP (RB now)
(ADJP
(NP
(QP (JJR more) (IN than) (CD 100))
(NNS years))
(JJ old))))
The English tag PRN (parenthetical) is used to tag the parenthetical contents
which are usually like sub-sentence and there is no such tag in the French phrase
tagset, so PRN is also considered as under the category of sub-sentence in the
universal tags, for instance:
(PRN (, ,)
(S
(NP (PRP it))
(VP (VBZ seems)))
(, ,))
(PRN (-LRB- -LRB-)
(NP
(NP (CD 204) (NNS boys))
(CC and)
(NP (CD 172) (NNS girls)))
(PP (IN from)
(NP (CD 13) (NNS schools)))
(-RRB- -RRB-))
The universal phrase tag COP (coordinated phrase) covers: the French tag
COORD (coordinated phrases); the English tag UCP (unlike coordinated phrase,
labeling the coordinated phrases that belong to diﬀerent categories).
The universal phrase tag CONJP (conjunction phrase) covers: the English tag
CONJP (conjunction phrase). The French Treebank has no homologous tags that
have the similar function to the Penn Treebank phrase tag CONJP (conjunction
phrase), for instance:
(CONJP (RB as) (RB well) (IN as))
The universal phrase tag X (other phrases or unknown) covers: the English
tags X (unknown or uncertain), INTJ (interjection), LST (list marker).
3 Application in MT Evaluation
The conventional MT evaluation methods include the measuring of the close-
ness of word sequences between the translated documents (hypothesis transla-
tion) and the reference translations (usually oﬀered by professional translators),
such as the BLEU [8], NIST [9] and METEOR [10], and the measuring of the
editing distance from the hypothesis translation to the reference translations,
such as the TER [11]. However, the reference translations are usually expen-
sive. Furthermore, the reference translations are in fact not available when the

translation task contains a large amount of data. The phrase tags mapping de-
signed for French and English treebanks is employed to develop a novel evalu-
ation approach for French-to-English MT translation without using the money-
and-time-consuming reference translations. The evaluation is performed on the
French (source) and English (target) documents.
3.1 Designed Methods
We assume that the translated (targeted) sentence should have a similar set of
phrase categories with the source sentence. For instance, if the source French
sentence has an adjective phrase and a prepositional phrase then the targeted
English sentence is most likely to have an adjective phrase and a preposition-
al phrase, too. This design is inspired by the synonymous relation between the
source sentence and the target sentence. Admittedly, it is possible that we get two
sentences that have the similar set of phrase categories but talk about different
things. However, this evaluation approach is not designed for the general circum-
stance. This evaluation approach is designed and applied under the assumption
that the targeted sentences are indeed the translated sentences (usually by some
automatic MT systems) from the source document. Under this assumption, if the
translated sentence retains the complete meaning of the source sentence, then it
usually contains most (if not all) of the phrase categories in the source sentence.
Furthermore, the Spearman correlation coefficient will be employed in the ex-
periments to test this evaluation approach (to measure the correlation scores of
this method derived by comparing with the human judgments). Following the
common practice, human judgments are assumed as the Gold Standard [12][13].
Assume we have the following two sentences: sentence X is the source French
and sentence Y is the targeted English (translated by a MT system). Firstly,
the two sentences are parsed by the corresponding language parsers respectively
(this paper uses the Berkeley parsers [14]). Then the phrase tags of each sen-
tence are extracted from the parsing results. Finally, the quality estimation of
the translated sentence Y is conducted based on these two sequences of extracted
phrase tags. See Figure 1 and Figure 2 for explanation (PhS and UniPhs rep-
resent phrase sequence and the universal type of phrase sequence respectively).
To make it both concise and content-rich, the phrase tag at the second level of
each word and punctuation (counted from bottom up, namely the tokens in the
parsing tree, i.e. just the immediate level above the POS tag level) is extracted.
Thus, the extracted phrase sequence (PhS) has the same length as that of the
parsed sentence.
3.2 Designed Formula
The precision and recall values are conventionally regarded as the important
factors to reflect the quality of system performances in NLP literature. For in-
stance, the BLEU metric employs uniformly weighted 4-gram precision scores
and METEOR employs both unigram precision and unigram recall combined

Fig. 1. The parsing results of French and English sentences.
Fig. 2. The converting of extracted phrase tags into universal categories.

with synonym and stem dictionaries. This paper designs a formula that contain-
s weighted N2 gram precision, weighted N3 gram recall and N1 gram position
order factors. Different n-gram precisions and recalls are combined respectively
with weighted geometric means (using different weight parameters, different with
BLEU). The whole formula is grouped using the weighted harmonic mean. We
design a factor, the position difference, due to the fact that there is indeed word
order information within the phrase. However, the request on the sequence or-
der becomes weak at phrase levels (e.g. some languages have free phrase order).
According to the analysis above, the weight of position factor will be adjusted
to be lower than the weights of precision and recall.
HPPR = Har(wP sN1PsDif, wP rN2Pre, wRcN3Rec)
=
wP s + wP r + wRc
wP s
N1P sDif
+ wP r
N2P re + wRc
N3Rec
(1)
In the formula, Har means harmonic mean; the variables, N1PsDif, N2Pre
and N3Rec, represent N1 gram position difference factor, weighted N2 gram
precision and weighted N3 gram recall (all at corpus level) respectively. The pa-
rameters wP s, wP r and wRc are the weights of the corresponding three variables
N1PsDif, N2Pre and N3Rec. Let’s see the introduction of the sentence-level
N1PsDif firstly.
N1PsDif = e−N1P D
(2)
N1PD =
1
Lengthhyp
Lengthhyp
i=1
| PDi | (3)
| PDi |=| PsNhyp − MatchPsNsrc | (4)
The variables, N1PD, Lengthhyp and PDi, mean N1 gram position difference
value of a sentence, hypothesis sentence length and position difference value of
the ith tag in the hypothesis sentence respectively. The variables PsNhyp and
MatchPsNsrc mean the position number of a tag in the hypothesis sentence and
the corresponding position number of the matching tag in the source sentence.
If there is no successful match for the hypothesis tag, then:
| PDi |=| PsNhyp − 0 | (5)
To calculate the value of N1PsDif, we should first finish the N1 gram UniPhS
alignment from the hypothesis to the source sentence (direction fixed). In this
case, the n-gram alignment means that the candidate matches that also have
neighbor matches (neighbors are limited to the previous and following N1 tags)
will be assigned higher priority. If there is no surrounding match for all potential
tags then the nearest match operation will be selected as a backup choice. On
the other hand, if there are more than two candidates in the source sentence that
both have the neighbor matching with the hypothesis tag, then the candidate

Fig. 3. The N1 gram alignment algorithm.
that has the nearest match with the hypothesis tag will also be selected. The
ﬁnal result is restricted to one-to-one alignment. The alignment algorithm is
shown in Figure 3. The variable lh and ls represent the length of hypothesis
sentence and source sentence respectively. The tagx and tags
y mean the phrase
tag in hypothesis and source sentence respectively. In this paper, the N1gram is
performed using bigram (N1=2). After the alignment, an example for the N1PD
calculation of a sentence is shown in Figure 4. Then the value of the N1 gram
Position diﬀerence factor N1PsDif at the corpus level is the arithmetical mean
of each sentence level score N1PsDifi.
N1PsDif =
1
n
n
i=1
N1PsDifi (6)
The values of the corpus-level weighted N2 gram precision N2Pre and weight-
ed N3 gram recall N3Rec are calculated as follows:
N2Pre = exp(
N2
n=1
wnlogPn) (7)
N3Rec = exp(
N3
n=1
wnlogRn) (8)

Fig. 4. The N1PD calculation example.
The variables Pn and Rn mean the n-gram precision and n-gram recall that
are calculated as follows:
Pn =
#matched ngram chunks
#ngram chunks of hypothesis corpus
(9)
Rn =
#matched ngram chunks
#ngram chunks of source corpus
(10)
The variable #matched ngram chunks represents the number of matched
n-gram chunks between the hypothesis and source corpus (sum of sentence-level
matches). Each n-gram chunk could be used only once, and an example is the
bigram chunks matching in Figure 5.
Fig. 5. Bigram chunks matching example.
To calculate n-gram precision and recall values, we should know that if they
are designed at the sentence level then some of the sentence-level bigram or tri-
gram precision and recall values will be 0 due to the fact that some hypothesis
sentences result in no bigram or trigram chunk matching with the source sen-
tences (it is meaningless when the variable of logarithmic function is less than
or equal to 0). According to the analysis above, after the sentence-level n-gram
chunks matching, we calculate the n-gram precision Pn and n-gram recall Rn
values at the corpus level as shown in the formula (this design is similar to BLEU;
BLEU calculates the corpus level n-gram precisions). Thus the geometric mean
of weighted N2 gram precision N2Pre and weighted N3 gram recall N3Rec will
also be at the corpus level.

4 Experiments
4.1 Evaluating method for evaluation
Spearman rank correlation coefficient (rs) is conventionally used as the evalu-
ating method for the MT evaluation metrics [12][13]. The evaluation result that
has a higher correlation score, namely higher correlation with human judgments,
will be regarded as having a better quality.
rsXY = 1 −
6
n
i=1 d2
i
n(n2 − 1)
(11)
We assume two sequence
−→
X = {x1, x2, ..., xn} and
−→
Y = {y1, y2, ..., yn}. First-
ly we get the corresponding rank sequences
ˆ−→
X = { ˆx1, ˆx2, ..., ˆxn} and
ˆ−→
Y =
{ ˆy1, ˆy2, ..., ˆyn}; then di is the difference-value between the coordinate variables,
i.e. ˆxi − ˆyi.
4.2 Experiments and Results
The corpora used in the experiments are from the international workshop of
statistical machine translation (WMT) which has been held annually by the
ACL’s special interest group of machine translation (SIGMT2
). To avoid the
overfitting problem, the WMT 2011 corpora are used as the development set (for
tuning the weights of factors in the formula to make the evaluation results close
to the human judgments). Then the WMT 2012 corpora are used as the testing
set with the formula that has the same parameters tuned in the development
stage.
There are 18 and 15 systems respectively in WMT 2011 and WMT 2012 pro-
ducing the French-to-English translation documents, each document containing
3003 sentences. Each year, there are hundreds of human annotators to evaluate
the MT system outputs, and the human judgments task usually costs hundred-
s of hours of labor. The human judgments are used to validate the automatic
metrics. The system-level Spearman correlation coefficient of the different eval-
uation results will be calculated as compared to the human judgments. The
state-of-the-art evaluation metrics BLEU (measuring the closeness between the
hypothesis and reference translations) and TER (measuring the editing distance)
are selected for the comparison with the designed model HPPR.
The values of N2 and N3 are both selected as 3 due to the fact that the 4-gram
chunk match usually results in 0 score. The tuned factor weights in the formula
are shown in Table 2. The experiment results on the development and testing
corpora are shown in Table 3 and Table 4 respectively (“Use Reference?” mean-
s whether this metric uses reference translations). The experiment results on
the development and testing corpora show that HPPR without using reference
translations has yielded promising correlation scores (0.63 and 0.59 respective-
ly) with human judgments even though lower than the reference-aware metrics
2
http://www.sigmt.org/

Table 2. Tuned factor weights
Items Weight parameters Ratio
N2Pre w1 : w2 : w3 8:1:1
N3Rec w1 : w2 : w3 1:1:8
HPPR wP s : wP r : wRc 1:8:1
Table 3. Development results on WMT 2011 corpus
Metrics Use Reference? Spearman rs
BLEU Yes 0.85
TER Yes 0.77
HPPR No 0.63
BLEU and TER. Results also specify that there is still potential to improve the
performances of all the three metrics, even though that the correlation scores
which are higher than 0.5 are already considered as strong correlation [15] as
shown in Table 5.
5 Discussion
To facilitate future research in multilingual or cross-lingual literature, this paper
designs a phrase tags mapping between the French Treebank and the English
Penn Treebank using 9 commonly used phrase categories. One of the potential
applications in the MT evaluation task is explored in the experiment section.
This novel evaluation model could spare the reference translation corpora that
are conventionally used in the MT evaluation tasks. This evaluation approach
is favorable especially for the circumstance where there are many translation
systems outputting a lot of translated documents and the reference translations
are very expensive or not easily available in time.
There are still some limitations in this work to be addressed in the future.
First, the designed universal phrase categories may not be able to cover all
the phrase tags of other language treebanks, so this tagset could be expanded
when necessary. Second, in the application of MT evaluation of the mapping
work, the designed formula contains the n-gram factors of position diﬀerence,
Table 4. Testing results on WMT 2012 corpus
Metrics Use Reference? Spearman rs
BLEU Yes 0.81
TER Yes 0.82
HPPR No 0.59

Table 5. Correlation scores analysis
Correlation Negative Positive
None -0.09 to 0 0 to 0.09
Small -0.3 to -0.1 0.1 to 0.3
Medium -0.5 to -0.3 0.3 to 0.5
Strong -1.0 to -0.5 0.5 to 1.0
precision and recall, which may not be sufficient or suitable for some of the other
language pairs, so different measuring factors should be added or switched when
facing new challenges.
The designed source codes of this paper can be freely downloaded
for research purpose, open source online. The source codes of HPPR mea-
suring algorithm and French-English phrase tagset mapping are available here
“https://github.com/aaronlifenghan/aaron-project-hppr”. 3
Acknowledgments. The authors wish to thank the anonymous reviewers for
many helpful comments. 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 Mr.
Yunsheng Xie who is a Master Candidate of Arts in English Studies from the
Department of English in the Faculty of Social Sciences and Humanities at the
University of Macau for his kind help in improving the language quality.
References
1. Marcus M., Beatrice Santorini and M.A. Marcinkiewicz. 1993. Building a large anno-
tated corpus of English: The Penn Treebank. In Computational Linguistics, volume
19, number 2, pp. 313-330.
2. Marcus Mitchell, Grace Kim, Mary Ann Marcinkiewicz, Robert MacIntyre, An-
n Bies, Mark Ferguson, Karen Katz and Britta Schasberger. 1994. THE PENN
TREEBANK: ANNOTATING PREDICATE ARGUMENT STRUCTURE, HU-
MAN LANGUAGE TECHNOLOGY: Proceedings of Workshop, Plainsboro, New
Jersey, March 8-11, pages 114-119. H94-1020.
3. Skut W., B. Krenn, T. Brants, and H. Uszkoreit. 1997. An annotation scheme for
free word order languages. In Proc. of ANLP, pages 88-95.
4. Abeillé A., L. Clément, and F. Toussenel. 2003. Building a Treebank for French.
Building and Using Parsed Corpora, In Abeillé (Abeillé, 2003), chapter 10 in ANNE
Abeillé, Treebanks. Kluwer Academic Publishers.
3
The final publication is available at
http://www.springer.com/computer/database+management+%26+information+retrieval/book/978-
3-642-40721-5

5. Chen K., C. Luo, M. Chang, F. Chen, C. Chen, C. Huang, and Z. Gao. 2003. Sinica
tree-bank: Design criteria, representational issues and implementation. In Abeill´e
(Abeill´e, 2003), chapter 13, pages 231–248.
6. Petrov Slav, Dipanjan Das, and Ryan McDonald. 2012. A Universal Part-of-Speech
Tagset, Proceedings of the Eight International Conference on Language Resources
and Evaluation (LREC’12), Istanbul, Turkey.
7. Bies, A., Mark Ferguson, Karen Katz, and Robert MacIntyre. 1995. Bracketing
Guidelines for Treebank II style Penn Treebank Project. Linguistic Data Consor-
tium.
8. Papineni Kishore, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. BLEU: a
method for automatic evaluation of machine translation. In Proceedings of the 40th
Annual Meeting on Association for Computational Linguistics (ACL ’02). Associa-
tion for Computational Linguistics, Stroudsburg, PA, USA, 311-318.
9. Doddington George. 2002. Automatic evaluation of machine translation quality us-
ing n-gram co-occurrence statistics. In Proceedings of the second international con-
ference on Human Language Technology Research (HLT ’02). Morgan Kaufmann
Publishers Inc., San Francisco, CA, USA, 138-145.
10. Banerjee Satanjeev and Alon Lavie. 2005. METEOR: An Automatic Metric for MT
Eval-uation with Improved Correlation with Human Judgments. In Proceedings of
the 43th An-nual Meeting of the Association of Computational Linguistics (ACL-
05), pages 65–72, Ann Arbor, US, June. Association of Computational Linguistics.
11. Snover Matthew, Bonnie J. Dorr, Richard Schwartz, Linnea Micciulla, and John
Makhoul. 2006. A Study of Translation Edit Rate with Targeted Human Annotation.
In Proceedings of the 7th Conference of the Association for Machine Translation in
the Americas (AMTA-06), pages 223–231, USA.
12. Callison-Burch, C., Koehn, P., Monz, C. and Zaidan, O. F. 2011. Findings of the
2011 Workshop on Statistical Machine Translation. In Proceedings of the Sixth
Workshop on Statistical Machine translation of the Association for Computational
Linguistics(ACL-WMT), pages 22-64, Edinburgh, Scotland, UK. Association for
Computational Linguistics.
13. Callison-Burch Chris, Philipp Koehn, Christof Monz, Matt Post, Radu Soricut, and
Lucia Specia. 2012. Findings of the 2012Workshop on Statistical Machine Transla-
tion. In Pro-ceedings of the Seventh Workshop on Statistical Machine Translation,
pages 10–51, Mon-treal, Canada. Association for Computational Linguistics.
14. Slav Petrov, Leon Barrett, Romain Thibaux, and Dan Klein. 2006. Learning accu-
rate, compact, and interpretable tree annotation. In Proceedings of the 21st Inter-
national Con-ference on Computational Linguistics and the 44th annual meeting of
the Association for Computational Linguistics (ACL-44). Association for Computa-
tional Linguistics, Strouds-burg, PA, USA, 433-440.
15. Cohen, J. 1988. Statistical power analysis for the behavioral sciences. 2nd ed.
Psychology Press.

GSCL2013.Phrase Tagset Mapping for French and English Treebanks and Its Application in Machine Translation Evaluation

More Related Content

Viewers also liked (7)

Similar to GSCL2013.Phrase Tagset Mapping for French and English Treebanks and Its Application in Machine Translation Evaluation (20)

More from Lifeng (Aaron) Han (20)

Recently uploaded (20)

GSCL2013.Phrase Tagset Mapping for French and English Treebanks and Its Application in Machine Translation Evaluation