Automatic Adaptation of Annotations

Wenbin Jiang∗
Chinese Academy of Sciences

Yajuan L ¨u∗
Chinese Academy of Sciences

Liang Huang∗∗
Queens College and Graduate Center,
The City University of New York

Qun Liu∗†
Dublin City University
Chinese Academy of Sciences

Manually annotated corpora are indispensable resources, yet for many annotation tasks, tel que
the creation of treebanks, there exist multiple corpora with different and incompatible annotation
guidelines. This leads to an inefﬁcient use of human expertise, but it could be remedied by
integrating knowledge across corpora with different annotation guidelines. In this article we
describe the problem of annotation adaptation and the intrinsic principles of the solutions, et
present a series of successively enhanced models that can automatically adapt the divergence
between different annotation formats.

We evaluate our algorithms on the tasks of Chinese word segmentation and dependency
parsing. For word segmentation, where there are no universal segmentation guidelines be-
cause of the lack of morphology in Chinese, we perform annotation adaptation from the much
larger People’s Daily corpus to the smaller but more popular Penn Chinese Treebank. Pour
dependency parsing, we perform annotation adaptation from the Penn Chinese Treebank to
a semantics-oriented Dependency Treebank, which is annotated using signiﬁcantly different
annotation guidelines. In both experiments, automatic annotation adaptation brings signiﬁcant
improvement, achieving state-of-the-art performance despite the use of purely local features in
entraînement.

∗ Key Laboratory of Intelligent Information Processing, Institute of Computing Technology, Chinese

Académie des Sciences, Non. 6 Kexueyuan South Road, Haidian District, P.O. Box 2704, Pékin 100190,
Chine. E-mail: {jiangwenbin, liuqun, lvyajuan}@ict.ac.cn.

∗∗ Department of Computer Science, Queens College / CUNY, 65-30 Kissena Blvd., Queens, New York 11367.

E-mail: liang.huang.sh@gmail.com.

† Centre for Next Generation Localisation, Faculty of Engineering and Computing, Dublin City University.

E-mail: qliu@computing.dcu.ie.

Submission received: 24 Avril 2013; revised version received: 6 Mars 2014; accepted for publication:
18 Avril 2014.

est ce que je:10.1162/COLI a 00210

D
o
w
n
o
un
d
e
d

F
r
o
m
h

t
t

:
/
/

d
je
r
e
c
t
.

je
t
.

e
d
toi
/
c
o

je
je
/

un
r
t
je
c
e
–
p
d

F
/

4
1
1
1
1
9
1
8
0
5
3
5
3
/
c
o

je
je

_
un
_
0
0
2
1
0
p
d

b
oui
g
toi
e
s
t

o
n
0
8
S
e
p
e
m
b
e
r
2
0
2
3

Computational Linguistics

Volume 41, Nombre 1

1. Introduction

Much of statistical NLP research relies on some sorts of manually annotated corpora
to train models, but annotated resources are extremely expensive to build, especially
on a large scale. The creation of treebanks is a prime example (Marcus, Santorini,
and Marcinkiewicz 1993). Cependant, the linguistic theories motivating these annotation
efforts are often heavily debated, and as a result there often exist multiple corpora
for the same task with vastly different and incompatible annotation philosophies. Pour
exemple, there are several treebanks for English, including the Chomskian-style Penn
Treebank (Marcus, Santorini, and Marcinkiewicz 1993), the HPSG LinGo Redwoods
Treebank (Oepen et al. 2002), and a smaller dependency treebank (Buchholz and Marsi
2006). From the perspective of resource accumulation, it seems a waste in human
efforts.1

A second, related problem is that the raw texts are also drawn from dif-
ferent domains, which for the above example range from ﬁnancial news (Penn
Treebank/Wall Street Journal) to transcribed dialog (LinGo). It would be nice if a
system could be automatically ported from one set of guidelines and/or domain to
another, in order to exploit a much larger data set. The second problem, domain
adaptation, is very well studied (par exemple., Blitzer, McDonald's, & Pereira 2006; Daum´e III
2007). This work focuses on the widely existing and equally important problem, un-
notation adaptation, in order to adapt the divergence between different annotation
guidelines and integrate linguistic knowledge in corpora with incongruent annotation
formats.

In this article, we describe the problem of annotation adaptation and the intrinsic
principles of the solutions, and present a series of successively improved concrete
models, the goal being to transfer the annotations of a corpus (source corpus) to the
annotation format of another corpus (target corpus). The transfer classiﬁer is the fun-
damental component for annotation adaptation algorithms. It learns correspondence
regularities between annotation guidelines from a parallel annotated corpus, which has
two kinds of annotations for the same data. In the simplest model (Model 1), the source
classiﬁer trained on the source corpus gives its predications to the transfer classiﬁer
trained on the parallel annotated corpus, so as to integrate the knowledge in the two
corpora. In a variant of the simplest model (Model 2), the transfer classiﬁer is used to
transform the annotations in the source corpus into the annotation format of the target
corpus; then the transformed source corpus and the target corpus are merged in order
to train a more accurate classiﬁer. Based on the second model, we ﬁnally develop an
optimized model (Model 3), where two optimization strategies, iterative training and
predict-self re-estimation, are integrated to further improve the efﬁciency of annotation
adaptation.

We experiment on Chinese word segmentation and dependency parsing to test
the efﬁcacy of our methods. For word segmentation, the problem of incompatible
annotation guidelines is one of the most glaring: No segmentation guideline has been
widely accepted due to the lack of a clear deﬁnition of Chinese word morphology.
For dependency parsing there also exist multiple disparate annotation guidelines. Pour

1 Different annotated corpora for the same task facilitate the comparison of linguistic theories. From this
perspective, having multiple standards is not necessarily a waste but rather a blessing, because it is a
necessary phase in coming to a consensus if there is one.

120

D
o
w
n
o
un
d
e
d

F
r
o
m
h

t
t

:
/
/

d
je
r
e
c
t
.

je
t
.

e
d
toi
/
c
o

je
je
/

un
r
t
je
c
e
–
p
d

F
/

4
1
1
1
1
9
1
8
0
5
3
5
3
/
c
o

je
je

_
un
_
0
0
2
1
0
p
d

b
oui
g
toi
e
s
t

o
n
0
8
S
e
p
e
m
b
e
r
2
0
2
3

Jiang et al.

Automatic Adaptation of Annotations

VV NR
NR
{1 B2 o3 Ú4 –5 u6
U.S. Vice-President visited China

n
ns
{1 B2 o3 Ú4 –5 u6
U.S. Vice

President visited-China

Figure 1
Incompatible word segmentation and POS tagging guidelines between CTB (upper) and PD
(below).

exemple, the dependency relations extracted from a constituency treebank follow syn-
tactic principles, whereas the semantic dependency treebank is annotated in a semantic
perspective.

The two corpora for word segmentation are the much larger People’s Daily cor-
pus (PD) (5.86M words) (Yu et al. 2001) and the smaller but more popular Penn
Chinese Treebank (CTB) (0.47M words) (Xue et al. 2005). They utilize very differ-
ent segmentation guidelines; Par exemple, as shown in Figure 1, PD breaks Vice-
President into two words and combines the phrase visited-China as a compound,
compared with the segmentation following the CTB annotation guideline. It is prefer-
able to transfer knowledge from PD to CTB because the latter also annotates tree
constructions, which are useful for downstream applications like parsing, summariza-
tion, and machine translation, yet it is much smaller in size. For dependency pars-
ing, we use the dependency treebank (DCTB) extracted from CTB according to the
rules of Yamada and Matsumoto (2003), and the Semantic Dependency Treebank
(SDT) built on a small part of the CTB text (Che et al. 2012). Compared with the
automatically extracted dependencies in DCTB, semantic dependencies in SDT re-
veal semantic relationships between words, rather than the syntactic relationships in
syntactic dependencies. Chiffre 2 shows an example. Experiments on both word seg-
mentation and dependency parsing show that annotation adaptation results in signiﬁ-
cant improvement over the baselines, and achieves the state-of-the-art with only local
features.

The rest of the article is organized as follows. Section 2 gives a description
of the problem of annotation adaptation. Section 3 brieﬂy introduces the tasks
of word segmentation and dependency parsing as well as their state-of-the-art
models. In Section 4 we ﬁrst describe the transfer classiﬁer that indicates the

(cid:1117)(cid:3373) (cid:4649) (cid:3910) (cid:5424)(cid:7022) (cid:6208)(cid:13593) (cid:7104)(cid:10086)

Figure 2
Incompatible dependency guidelines between SDT (top) and DCTB (bottom).

121

D
o
w
n
o
un
d
e
d

F
r
o
m
h

t
t

:
/
/

d
je
r
e
c
t
.

je
t
.

e
d
toi
/
c
o

je
je
/

un
r
t
je
c
e
–
p
d

F
/

4
1
1
1
1
9
1
8
0
5
3
5
3
/
c
o

je
je

_
un
_
0
0
2
1
0
p
d

b
oui
g
toi
e
s
t

o
n
0
8
S
e
p
e
m
b
e
r
2
0
2
3

Computational Linguistics

Volume 41, Number 1

intrinsic principles of annotation adaptation, and then depict the three successively
enhanced models for automatic adaptation of annotations. After the description of
experimental results in Section 5 and the discussion of application scenarios in Sec-
tion 6, we give a brief review of related work in Section 7, drawing conclusions in
Section 8.

2. Automatic Annotation Adaptation

We deﬁne annotation adaptation as a task aimed at automatically adapting the
divergence between different annotation guidelines. Statistical models can be de-
two annotation guidelines in order to trans-
signed to learn the relevance of
form a corpus from one annotation guideline to another. From this point of
voir, annotation adaptation can be seen as a special case of transfer learning.
Through annotation adaptation, the linguistic knowledge in different corpora is
integrated, resulting in enhanced NLP systems without complicated models and
features.

Much research has considered the problem of domain adaptation (Blitzer,
McDonald's, and Pereira 2006; Daum´e III 2007), which also can be seen as a special case
of transfer learning. It aims to adapt models trained in one domain (par exemple., chemistry) à
work well in other domains (par exemple., medicine). Despite superﬁcial similarities between
domain adaptation and annotation adaptation, we argue that the underlying problems
are quite different. Domain adaptation assumes that the labeling guidelines are
preserved between the two domains—for example, an adjective is always labeled
as JJ regardless of whether it is from the Wall Street Journal (WSJ) or a biomedical
text, and only the distributions are different—for example, the word control is most
likely a verb in WSJ but often a noun in biomedical texts (as in control experiment).
Annotation adaptation, cependant, tackles the problem where the guideline itself
is changed, Par exemple, one treebank might distinguish between transitive and
intransitive verbs, while merging the different noun types (NN, NNS, etc.), or one
treebank (PTB) might be much ﬂatter than the other (LinGo), not to mention the
fundamental disparities between their underlying linguistic representations (CFG vs.
HPSG).

A more formal description will allow us to make these claims more precise. Let X
be the data and Y be the annotation. Annotation adaptation can be understood as a
change of P(Oui) due to the change in annotation guidelines while P(X) remains constant.
Through annotation adaptation, we want to change the annotations of the data from one
guideline to another, leaving the data itself unchanged. Cependant, in domain adaptation,
P.(X) changes, but P(Oui) is assumed to be constant. The word assumed means that the
distributions P(Oui, X) and P(Oui|X) are actually changed because P(X) is changed. Domain
adaptation aims to make the model better adapt to a different domain with the same
annotation guidelines.

According to this analysis, annotation adaptation seems more motivated from a
linguistic (rather than statistical) point of view, and tackles a serious problem fun-
damentally different from domain adaptation, which is also a serious problem (souvent
leading to >10% loss in accuracy). More interestingly, annotation adaptation, without
any assumptions about distributions, can be simultaneously applied to both domain
and annotation adaptation problems, which is very appealing in practice because the
latter problem often implies the former.

122

D
o
w
n
o
un
d
e
d

F
r
o
m
h

t
t

:
/
/

d
je
r
e
c
t
.

je
t
.

e
d
toi
/
c
o

je
je
/

un
r
t
je
c
e
–
p
d

F
/

4
1
1
1
1
9
1
8
0
5
3
5
3
/
c
o

je
je

_
un
_
0
0
2
1
0
p
d

b
oui
g
toi
e
s
t

o
n
0
8
S
e
p
e
m
b
e
r
2
0
2
3

Jiang et al.

Automatic Adaptation of Annotations

3. Études de cas: Word Segmentation and Dependency Parsing

3.1 Word Segmentation and Character Classiﬁcation Method

In many Asian languages there are no explicit word boundaries, thus word segmen-
tation is a fundamental task for the processing and understanding of these languages.
Given a sentence as a sequence of n characters:

x = x1 x2 .. xn

where xi is a character, word segmentation aims to split the sequence into m(≤ n)
mots:

x1:e1 xe1+1:e2

.. xem−1+1:em

where each subsequence xi:j
xi to xj.

indicates a Chinese word spanning from characters

Word segmentation can be formalized as a sequence labeling problem (Xue and
Shen 2003), where each character in the sentence is given a boundary tag representing
its position in a word. Following Ng and Low (2004), joint word segmentation and part-
of-speech (POS) tagging can also be solved using a character classiﬁcation approach by
extending boundary tags to include POS information. For word segmentation we adopt
the four boundary tags of Ng and Low (2004), B, M., E, et S, where B, M., and E mean
the beginning, the middle, and the end of a word, respectivement, and S indicates a single-
character word. The word segmentation result can be generated by splitting the labeled
character sequence into subsequences of pattern S or BM∗E, indicating single-character
words or multi-character words, respectively.

Given the character sequence x, the decoder ﬁnds the output ˜y that maximizes the

score function:

˜y = argmax

F(X, oui) · w

oui

= argmax

oui

X
xi∈x,yi∈y

F(xi, yi) · w

(1)

Where function f maps (X, oui) into a feature vector, w is the parameter vector generated
by the training algorithm, and f(X, oui) · w is the inner product of f(X, oui) and w. Le
score of the sentence is further factorized into each character, where yi is the character
classiﬁcation label of character xi.

The training procedure of perceptron learns a discriminative model mapping from
the inputs x to the outputs y. Algorithm 1 shows the perceptron algorithm for tuning the
parameter w. The “averaged parameters” technology (Collins 2002) is used for better
performance. The feature templates of the classiﬁer is shown in Table 1. The function
Pu(·) returns true for a punctuation character and false for others; the function T(·) clas-
siﬁes a character into four types: number, date, English letter, et d'autres, corresponding
to function values 1, 2, 3, et 4, respectively.

3.2 Dependency Parsing and Spanning Tree Method

Dependency parsing aims to link each word to its arguments so as to form a directed
graph spanning the whole sentence. Normally, the directed graph is restricted to a

123

D
o
w
n
o
un
d
e
d

F
r
o
m
h

t
t

:
/
/

d
je
r
e
c
t
.

je
t
.

e
d
toi
/
c
o

je
je
/

un
r
t
je
c
e
–
p
d

F
/

4
1
1
1
1
9
1
8
0
5
3
5
3
/
c
o

je
je

_
un
_
0
0
2
1
0
p
d

b
oui
g
toi
e
s
t

o
n
0
8
S
e
p
e
m
b
e
r
2
0
2
3

Computational Linguistics

Volume 41, Number 1

Algorithm 1 Perceptron training algorithm.
1: Input: Training set C
2: w ← 0
3: for t ← 1 .. T do
4:
5:

pour (X, oui) ∈ C do

˜z ← argmaxz f(X, z) · w
if ˜z 6= y then

w ← w + F(X, oui) − f(X, ˜z)

7:
8:

end if

⊲ T iterations

⊲ update the parameters

D
o
w
n
o
un
d
e
d

F
r
o
m
h

t
t

:
/
/

d
je
r
e
c
t
.

je
t
.

e
d
toi
/
c
o

je
je
/

un
r
t
je
c
e
–
p
d

F
/

4
1
1
1
1
9
1
8
0
5
3
5
3
/
c
o

je
je

_
un
_
0
0
2
1
0
p
d

b
oui
g
toi
e
s
t

o
n
0
8
S
e
p
e
m
b
e
r
2
0
2
3

end for

9:
10: end for
11: Output: Parameters w

dependency tree where each word depends on exactly one parent, and all words ﬁnd
their parents. Given a sentence as a sequence n words:

x = x1 x2 .. xn

dependency parsing ﬁnds a dependency tree y, où (je, j) ∈ y is an edge from the head
word xi to the modiﬁer word xj. The root r ∈ x in the tree y has no head word, and each
of the other words, j(j ∈ x and j 6= r), depends on a head word i(i ∈ x and i 6= j).

For many languages, the dependency structures are supposed to be projective. Si
xj is dependent on xi, then all the words between i and j must be directly or indirectly
dependent on xi. Donc, if we put the words in their linear order, preceded by the
root, all edges can be drawn above the words without crossing. We follow this constraint
because the dependency treebanks in our experiments are projective.

Following the edge-based factorization method (Eisner 1996), the score of a de-
pendency tree can be factorized into the dependency edges in the tree. The spanning

Table 1
Feature templates and instances for character classiﬁcation-based word segmentation model.
C0 denotes the current character, and C−i/Ci denote the ith character to the left/right of C0.
Suppose we are considering the third character “总” in “美-副-总-统-访-华”.

Type

Templates

Instances

n-gram

C−2
C−1
C0
C1
C2
C−2C−1
C−1C0
C0C1
C1C2
C−1C1

C−2=美
C−1=副
C0=总
C1=统
C2=访
C−2C−1=美副
C−1C0=副总
C0C1=总统
C1C2=统访
C−1C1=副统

fonction

Pu(C0)
T(C−2:2)

Pu(C0)=true
T(C−2:2)= 44444

124

Jiang et al.

Automatic Adaptation of Annotations

tree method (McDonald's, Crammer, and Pereira 2005) factorizes the score of the tree as
the sum of the scores of all its edges, and the score of an edge is deﬁned as the inner
product of the feature vector f and the weight vector w. Given a sentence x, the parsing
procedure searches for the candidate dependency tree with the maximum score:

˜y = argmax

F(X, oui) · w

oui

= argmax

oui

X
(je,j)∈y

F(je, j) · w

(2)

The averaged perceptron algorithm is used again to train the parameter vector. UN
bottom–up dynamic programming algorithm is designed to search for the candidate
parse with the maximum score as shown in Algorithm 2, where V[je, j] contains the
candidate dependency fragments of the span [je, j]. The feature templates are similar
to those of the ﬁrst-ordered MST model (McDonald's, Crammer, and Pereira 2005). Chaque
feature is composed of some words and POS tags surround word i and/or word j, comme
well as an optional distance representation between the two words. Tableau 2 shows the
feature templates without distance representations.

4. Models for Automatic Annotation Adaptation

Dans cette section, we present a series of discriminative learning algorithms for the auto-
matic adaptation of annotation guidelines. To facilitate the description of the algorithms,
several shortened forms are adopted for convenience of description. We use source
corpus to denote the corpus with the annotation guideline that we do not require, lequel
is of course the source side of adaptation, and target corpus denotes the corpus with
the desired guideline. Correspondingly, the annotation guidelines of the two corpora
are denoted as source guideline and target guideline, and the classiﬁers following the
two annotation guidelines are respectively named as source classiﬁer and target clas-
siﬁer. Given a parallel annotated corpus, c'est, a corpus labeled with two annotation

Algorithm 2 Dependency parsing algorithm.

4:
5:

7:
8:

1: Input: sentence x to be parsed
2: pour [je, j] ⊆ [1, |X|] in topological order do
3:

buf ← ∅
for k ← i..j − 1 faire

for l ∈ V[je, k] and r ∈ V[k + 1, j] faire
insert DERIV(je, r) into buf
insert DERIV(r, je) into buf

end for

end for
V[je, j] ← best K in buf

10:
11: end for
12: Output: the best of V[1, |X|]
13: function DERIV(p, c)
14:
15: end function

return p ∪ c ∪ {(p · root, c · root)}

⊲ all partitions

⊲ new derivation

125

D
o
w
n
o
un
d
e
d

F
r
o
m
h

t
t

:
/
/

d
je
r
e
c
t
.

je
t
.

e
d
toi
/
c
o

je
je
/

un
r
t
je
c
e
–
p
d

F
/

4
1
1
1
1
9
1
8
0
5
3
5
3
/
c
o

je
je

_
un
_
0
0
2
1
0
p
d

b
oui
g
toi
e
s
t

o
n
0
8
S
e
p
e
m
b
e
r
2
0
2
3

Computational Linguistics

Volume 41, Nombre 1

Tableau 2
Feature templates and instances for dependency parsing. Wi/Pi denotes the word/POS of the
hypothetic head, Wj/Pj denotes the word/POS of the hypothetic modiﬁer, and Px+/−1 denotes
the POS to the right/left of the x-th token. Suppose we are considering the words “外” (i = 2)
and “开放” (j = 3) in “中国/NR-对/P-外/NN-开放/VV-成绩/NN-斐然/VV”.

Type

Templates

Instances

unigram WiPi

bigram

contexte

Wi
Pi
WjPj
Wj
PJ

WiPiWjPj
WiWjPj
PiWjPj
WiPiWj
WiPiPj
WiWj
WiPj
PiWj
PiPj

WiPi=外◦NN
Wi=外
Pi=NN
WjPj=开放◦VV
Wj=开放
Pj=VV

WiPiWjPj=外◦NN◦开放◦VV
WiWjPj=外◦开放◦VV
PiWjPj=NN◦开放◦VV
WiPiWj=外◦NN◦开放
WiPiPj=外◦NN◦VV
WiWj=外◦开放
WiPj=外◦VV
PiWj=NN◦开放
PiPj=NN◦VV

PiPi+1Pj−1Pj
Pi−1PiPj−1Pj
PiPi+1PjPj+1
Pi−1PiPjPj+1
Pi−1PiPj−1
Pi−1PiPj+1
PiPi+1Pj−1
PiPi+1Pj+1
Pi−1Pj−1Pj
Pi−1PjPj+1
Pi+1Pj−1Pj
Pi+1PjPj+1
PiPj−1Pj
PiPjPj+1
Pi−1PiPj
PiPi+1Pj

PiPi+1Pj−1Pj=NN◦VV◦NN◦VV
Pi−1PiPj−1Pj=P◦NN◦NN◦VV
PiPi+1PjPj+1=NN◦VV◦VV◦NN
Pi−1PiPjPj+1=P◦NN◦VV◦NN
Pi−1PiPj−1=P◦NN◦NN
Pi−1PiPj+1=P◦NN◦NN
PiPi+1Pj−1=NN◦VV◦NN
PiPi+1Pj+1=NN◦VV◦NN
Pi−1Pj−1Pj=P◦NN◦VV
Pi−1PjPj+1=P◦VV◦NN
Pi+1Pj−1Pj=VV◦NN◦VV
Pi+1PjPj+1=VV◦VV◦NN
PiPj−1Pj=NN◦NN◦VV
PiPjPj+1=NN◦VV◦NN
Pi−1PiPj=P◦NN◦VV
PiPi+1Pj=NN◦VV◦VV

guidelines, a transfer classiﬁer can be trained to capture the regularity of transforma-
tion from the source annotation to the target annotation. The classiﬁers mentioned
here are normal discriminative classiﬁers that take a set of features as input and give
a classiﬁcation label as output. For the POS tagging problem, the classiﬁcation label
is a POS tag, and for the parsing task, the classiﬁcation label is a dependency edge, un
constituency span, or a shift-reduce action.

The parallel annotated corpus is the knowledge source of annotation adaptation.
The annotation quality and data size of the parallel annotated corpus determine the
accuracy of the transfer classiﬁer. Such a corpus is difﬁcult to build manually, bien que
we can generate a noisy one automatically from the source corpus and the target corpus.
Par exemple, we can apply the source classiﬁer on the target corpus, thus generating

126

D
o
w
n
o
un
d
e
d

F
r
o
m
h

t
t

:
/
/

d
je
r
e
c
t
.

je
t
.

e
d
toi
/
c
o

je
je
/

un
r
t
je
c
e
–
p
d

F
/

4
1
1
1
1
9
1
8
0
5
3
5
3
/
c
o

je
je

_
un
_
0
0
2
1
0
p
d

b
oui
g
toi
e
s
t

o
n
0
8
S
e
p
e
m
b
e
r
2
0
2
3

Jiang et al.

Automatic Adaptation of Annotations

a parallel annotated corpus with noisy source annotations and accurate target anno-
tations. The training procedure of the transfer classiﬁer predicts the target annotations
with guiding features extracted from the source annotations. This approach can alleviate
the effect of the noise in source annotations, and learn annotation adaptation regularities
accurately. By reducing the noise in the automatically generated parallel annotated
corpus, a higher accuracy of annotation adaptation can be achieved.

In the following sections, we ﬁrst describe the transfer classiﬁer that reveals the
intrinsic principles of annotation adaptation, then describe a series of successively
enhanced models that are developed from our previous investigation (Jiang, Huang,
and Liu 2009; Jiang et al. 2012). In the simplest model (Model 1), two classiﬁers, a source
classiﬁer and a transfer classiﬁer, are used in a cascade. The classiﬁcation results of the
lower source classiﬁer provide additional guiding features to the upper transfer classi-
ﬁer, yielding an improved classiﬁcation result. A variant of the ﬁrst model (Model 2)
uses the transfer classiﬁer to transform the source corpus from source guideline to
target guideline ﬁrst, then merges the transformed source corpus into the target corpus
in order to train an improved target classiﬁer on the enlarged corpus. An optimized
model (Model 3) is further proposed based on Model 2. Two optimization strategies,
iterative training and predict-self re-estimation, are adopted to improve the efﬁciency
of annotation adaptation, in order to fully utilize the knowledge in heterogeneous
corpora.

4.1 Transfer Classiﬁer

In order to learn the regularity of the adaptation from one annotation guideline to
another, a parallel annotated corpus is needed to train the transfer classiﬁer. The parallel
annotated corpus is a corpus with two different annotation guidelines, the source guide-
line and the target guideline. With the target annotation labels as learning objectives
and the source annotation labels as guiding information, the transfer classiﬁer learns
the statistical regularity of the adaptation from the source annotations to the target
annotations.

The training procedure of the transfer classiﬁer is analogous to the training of a
normal classiﬁer except for the introduction of additional guiding features. For word
segmentation, the most intuitive guiding feature is the source annotation label itself. Pour
dependency parsing, an effective guiding feature is the dependency path between the
hypothetic head and modiﬁer, as shown in Figure 3. Cependant, our effort is not limited
to this, and more special features are introduced: A classiﬁcation label or dependency
path is attached to each feature of the baseline classiﬁer to generate combined guiding
features. This is similar to the feature design in discriminative dependency parsing
(McDonald's, Crammer, and Pereira 2005; McDonald and Pereira 2006), where the basic
features, composed of words and POSs in the context, are also conjoined with link
direction and distance in order to generate more special features.

Tableau 3 shows an example of guide features (as well as baseline features) for word
segmentation, where “α = B” indicates that the source classiﬁcation label of the current
character is B, demarcating the beginning of a word. The combination strategy derives
a series of speciﬁc features, helping the transfer classiﬁer to produce more precise clas-
siﬁcations. The parameter-tuning procedure of the transfer classiﬁer will automatically
learn the regularity of using the source annotations to guide the classiﬁcation decision.
In decoding, if the current character shares some basic features in Table 3 and it is
classiﬁed as B in the source annotation, then the transfer classiﬁer will probably classify
it as M.

127

D
o
w
n
o
un
d
e
d

F
r
o
m
h

t
t

:
/
/

d
je
r
e
c
t
.

je
t
.

e
d
toi
/
c
o

je
je
/

un
r
t
je
c
e
–
p
d

F
/

4
1
1
1
1
9
1
8
0
5
3
5
3
/
c
o

je
je

_
un
_
0
0
2
1
0
p
d

b
oui
g
toi
e
s
t

o
n
0
8
S
e
p
e
m
b
e
r
2
0
2
3

Computational Linguistics

Volume 41, Nombre 1

(cid:1117)(cid:3373) (cid:4649) (cid:3910) (cid:5424)(cid:7022) (cid:6208)(cid:13593) (cid:7104)(cid:10086)

( (cid:6208)(cid:13593)(cid:3910)(cid:68)
,

(cid:32))

Chiffre 3
The guiding feature for dependency parsing, where α(je, j) denotes the dependency path between
i and j. In this example, j is a son of the great-grandfather of i.

En outre, the original features used in the normal classiﬁer are also used in order
to leverage the knowledge from the target annotation of the parallel annotated corpus,
and the training procedure of the transfer classiﬁer also learns the relative weights
between the guiding features and the original features. Donc, the knowledge from
both the source annotation and the target annotation are automatically integrated
ensemble, and higher and more stable prediction accuracy can be achieved.

Tableau 3
Feature templates and corresponding instances for the transfer classiﬁer. Suppose we are
considering the third character “总” in “美-副-总-统-访-华”. α = B indicates that the classiﬁcation
label given by the source classiﬁer is B.

Type

Templates

Instances

Original C−2
C−1
C0
C1
C2
C−2C−1
C−1C0
C0C1
C1C2
C−1C1
Pu(C0)
T(C−2:2)

Guiding α

C−2=美
C−1=副
C0=总
C1=统
C2=访
C−2C−1=美副
C−1C0=副总
C0C1=总统
C1C2=统访
C−1C1=副统
Pu(C0)=true
T(C−2:2)= 44444

α=B
α=B ◦ C−2=美
α=B ◦ C−1=副
α=B ◦ C0=总
α=B ◦ C1=统
α=B ◦ C2=访
α=B ◦ C−2C−1=美副
α=B ◦ C−1C0=副总
α=B ◦ C0C1=总统
α=B ◦ C1C2=统访
α=B ◦ C−1C1=副统
α=B ◦ Pu(C0)=true

α ◦ C−2
α ◦ C−1
α ◦ C0
α ◦ C1
α ◦ C2
α ◦ C−2C−1
α ◦ C−1C0
α ◦ C0C1
α ◦ C1C2
α ◦ C−1C1
α ◦ Pu(C0)
α ◦ T(C−2:2) α=B ◦ T(C−2:2)= 44444

128

D
o
w
n
o
un
d
e
d

F
r
o
m
h

t
t

:
/
/

d
je
r
e
c
t
.

je
t
.

e
d
toi
/
c
o

je
je
/

un
r
t
je
c
e
–
p
d

F
/

4
1
1
1
1
9
1
8
0
5
3
5
3
/
c
o

je
je

_
un
_
0
0
2
1
0
p
d

b
oui
g
toi
e
s
t

o
n
0
8
S
e
p
e
m
b
e
r
2
0
2
3

Jiang et al.

Automatic Adaptation of Annotations

source corpus

train with
normal features

target corpus

source classiﬁer

transformed
target corpus

train with
guiding features

transfer classiﬁer

Chiffre 4
The pipeline for training of Model 1.

4.2 Model 1

The most intuitive model for annotation adaptation uses two cascaded classiﬁers, le
source classiﬁer and the transfer classiﬁer, to integrate the knowledge in corpora with
different annotation guidelines. In the training procedure, a source classiﬁer is trained
on the source corpus and is used to process the target corpus, generating a parallel
annotated corpus (albeit a noisy one). Alors, the transfer classiﬁer is trained on the
parallel annotated corpus,with the target annotations as the classiﬁcation labels, et
the source annotation as guiding information. Chiffre 4 depicts the training pipeline. Le
best training iterations for the source classiﬁer and the transfer classiﬁer are determined
on the development sets of the source corpus and target corpus.

In the decoding procedure, a sequence of characters (for word segmentation) ou
mots (for dependency parsing) is input into the source classiﬁer to obtain a classiﬁca-
tion result under the source guideline; then it is input into the transfer classiﬁer with
this classiﬁcation result as the guiding information to get the ﬁnal result following the
target guideline. This coincides with the stacking method for combining dependency
parsers (Martins et al. 2008; Nivre and McDonald 2008), and is also similar to the Pred
baseline for domain adaptation in Daum´e et al. (Daum´e III and Marcu 2006; Daum´e III
2007). Chiffre 5 shows the pipeline for decoding.

raw sentence

source classiﬁer

result with
source guideline

transfer classiﬁer

result with
target guideline

Chiffre 5
The pipeline for decoding of Model 1.

129

D
o
w
n
o
un
d
e
d

F
r
o
m
h

t
t

:
/
/

d
je
r
e
c
t
.

je
t
.

e
d
toi
/
c
o

je
je
/

un
r
t
je
c
e
–
p
d

F
/

4
1
1
1
1
9
1
8
0
5
3
5
3
/
c
o

je
je

_
un
_
0
0
2
1
0
p
d

b
oui
g
toi
e
s
t

o
n
0
8
S
e
p
e
m
b
e
r
2
0
2
3

Computational Linguistics

Volume 41, Nombre 1

4.3 Model 2

The previous model has a drawback: It has to cascade two classiﬁers in decoding to
integrate the knowledge in two corpora, which seriously degrades the processing speed.
Here we describe a variant of the previous model, aiming at automatic transforma-
tion (rather than integration as in Model 1) between annotation guidelines of human-
annotated corpora. The source classiﬁer and the transfer classiﬁer are trained in the
same way as in the previous model. The transfer classiﬁer is used to process the source
corpus, with the source annotations as guiding information, so as to relabel the source
corpus with the target annotation guideline. By merging the target corpus and the
transformed source corpus for the training of the ﬁnal classiﬁer, improved classiﬁcation
accuracy can be achieved.

From this point on we describe the pipelines of annotation transformation in
pseudo-codes for simplicity and convenience of extensions. Algorithm 3 shows the
overall training algorithm for the variant model. Cs and Ct denote the source corpus
and the target corpus. Ms and Ms→t denote the source classiﬁer and the transfer
classiﬁer. Cq
p denotes the p corpus relabeled in q annotation guideline; Par exemple, Ct
s
is a corpus that labels the text of the source corpus with the target guideline. Func-
tions TRAIN and TRANSTRAIN train the source classiﬁer and the transfer classiﬁer,
respectivement, both invoking the perceptron algorithm, but with different feature sets.
Functions ANNOTATE and TRANSANNOTATE call the function DECODE with different
models (source/transfer classiﬁers), feature functions (without/with guiding features),
and inputs (raw/source-annotated sentences).

In the algorithm the parameters corresponding to development sets are omitted for
simplicity. Compared to the online knowledge integration methodology of the previous
model, annotation transformation leads to improved performance in an ofﬂine manner
by integrating corpora before the training procedure. This approach enables processing
speeds several times faster than the cascaded classiﬁers in the previous model. It also
has another advantage in that we can integrate the knowledge in more than two corpora
without slowing down the processing of the ﬁnal classiﬁer.

4.4 Model 3

The training of the transfer classiﬁer is based on an automatically generated (plutôt
than a gold standard) parallel annotated corpus, where the source annotations are

Algorithm 3 Baseline annotation adaptation.

D
o
w
n
o
un
d
e
d

F
r
o
m
h

t
t

:
/
/

d
je
r
e
c
t
.

je
t
.

e
d
toi
/
c
o

je
je
/

un
r
t
je
c
e
–
p
d

F
/

4
1
1
1
1
9
1
8
0
5
3
5
3
/
c
o

je
je

_
un
_
0
0
2
1
0
p
d

b
oui
g
toi
e
s
t

o
n
0
8
S
e
p
e
m
b
e
r
2
0
2
3

⊲ source classiﬁer

⊲ transfer classiﬁer

⊲ integrated corpus with target guideline

t , Ct)

1: function ANNOTRANS(Cs, Ct)
2: Ms ← TRAIN(Cs)
Cs
t ← ANNOTATE(MS, Ct)
3:
4: Ms→t ← TRANSTRAIN(Cs
5:

Ct
s ← TRANSANNOTATE(Ms→t, Cs)
Ct
∗ ← Ct
s ∪ Ct
6:
return Ct
7:
∗
8: end function
9: function DECODE(M., ΦΦΦ, X)
10:
11: end function

return argmaxy∈GEN(X) S(oui|M., ΦΦΦ, X)

130

Jiang et al.

Automatic Adaptation of Annotations

provided by the source classiﬁer. Donc, the performance of annotation transfor-
mation is correspondingly determined by the accuracy of the source classiﬁer, and we
can generate a more accurate parallel annotated corpus for better annotation adaptation
if an improved source classiﬁer can be obtained. Based on Model 2, two optimization
strategies—iterative bidirectional training and predict-self hypothesis—are introduced
to optimize the parallel annotated corpora for better annotation adaptation.

We ﬁrst use an iterative training procedure to gradually improve the transformation
accuracy by iteratively optimizing the parallel annotated corpora. In each training
itération, both source-to-target and target-to-source annotation transformations are per-
formed, and the transformed corpora are used to provide better annotations for the
parallel annotated corpora of the next iteration. Then in the new iteration, the better
parallel annotated corpora will result in more accurate transfer classiﬁers, resulting in
the generation of better transformed corpora.

Algorithm 4 shows the overall procedure of the iterative training method. Le
loop of lines 6–13 iteratively performs source-to-target and target-to-source annota-
tion transformations. The source annotations of the parallelly annotated corpora, Cs
t
and Ct
s, are initialized by applying the source and target classiﬁers on the target and
source corpora, respectivement (lines 2–5). In each training iteration, the transfer classiﬁers
are trained on the current parallel annotated corpora (lines 7–8); they are used to
produce the transformed corpora (lines 9–10), which provide better annotations for
the parallel annotated corpora of the next iteration. The iterative training terminates
when the performance of the classiﬁer trained on the merged corpus Ct
s ∪ Ct converges
(line 13).

The discriminative training of TRANSTRAIN predicts the target annotations with the
guidance of source annotations. In the ﬁrst iteration, the transformed corpora generated
by the transfer classiﬁers are better than the initial ones generated by the source and tar-
get classiﬁers, due to the assistance of the guiding features. In the following iterations,

D
o
w
n
o
un
d
e
d

F
r
o
m
h

t
t

:
/
/

d
je
r
e
c
t
.

je
t
.

e
d
toi
/
c
o

je
je
/

un
r
t
je
c
e
–
p
d

F
/

4
1
1
1
1
9
1
8
0
5
3
5
3
/
c
o

je
je

_
un
_
0
0
2
1
0
p
d

b
oui
g
toi
e
s
t

o
n
0
8
S
e
p
e
m
b
e
r
2
0
2
3

⊲ source classiﬁer

⊲ target classiﬁer

⊲ source-to-target transfer classiﬁer
⊲ target-to-source transfer classiﬁer

Algorithm 4 Iterative annotation transformation.

Cs
t ← ANNOTATE(MS, Ct)

1: function ITERANNOTRANS(Cs, Ct)
2: Ms ← TRAIN(Cs)
3:
4: Mt ← TRAIN(Ct)
5:
6:

Ct
s ← ANNOTATE(Mt, Cs)
repeat

7:
8:

10:
11:

12:

t , Ct)
s, Cs)

Ms→t ← TRANSTRAIN(Cs
Mt→s ← TRANSTRAIN(Ct
Ct
s ← TRANSANNOTATE(Ms→t, Cs)
Cs
t ← TRANSANNOTATE(Mt→s, Ct)
∗ ← Ct
Ct
s ∪ Ct
M∗ ← TRAIN(Ct
∗)

⊲ enhanced classiﬁer trained on merged corpus

until EVAL(M∗) converges
13:
return Ct
14:
∗
15: end function
16: function DECODE(M., ΦΦΦ, X)
17:
18: end function

return argmaxy∈GEN(X) S(oui|M., ΦΦΦ, X)

131

Computational Linguistics

Volume 41, Nombre 1

the transformed corpora provide better annotations for the parallel annotated corpora
of the subsequent iteration; the transformation accuracy will improve gradually along
with the optimization of the parallel annotated corpora until convergence.

The predict-self hypothesis is introduced to improve the transformation accuracy
from another perspective. This hypothesis is implicit in many unsupervised learning
approaches, such as Markov random ﬁeld; it has also been successfully used by Daum´e
III (2009) in unsupervised dependency parsing. The basic idea of predict-self is, if a
prediction is a better candidate for an input, it would be easier to convert it back to the
original input by a reverse procedure. If applied to annotation transformation, predict-
self indicates that a better transformation candidate following the target guideline can
be more easily transformed back to the original form following the source guideline.

The most intuitive strategy to introduce the predict-self methodology into annota-
tion transformation is using a reversed annotation transformation procedure to ﬁlter
out unreliable predictions of the previous transformation. In detail, a source-to-target
annotation transformation procedure is performed on the source corpus to obtain a pre-
diction that follows the target guideline; then a second, target-to-source transformation
procedure is performed on this prediction result to check whether it can be transformed
back to the original source annotation. The source-to-target prediction results that fail
in this reverse-veriﬁcation step are discarded, so this strategy can be called predict-self
ﬁltering.

A more sophisticated strategy can be called predict-self re-estimation. Instead of
using the reversed transformation procedure for ﬁltering, the re-estimation strategy
integrates the scores given by the source-to-target and target-to-source annotation
transformation models when evaluating the transformation candidates. By properly
tuning the relative weights of the two transformation directions, better transformation
performance is achieved. The scores of the two transformation models are weighted,
integrated in a log-linear manner:

S+(oui|Ms→t, Mt→s, ΦΦΦ, X) = (1 − λ) × S(oui|Ms→t, ΦΦΦ, X) + λ × S(X|Mt→s, ΦΦΦ, oui)

(3)

The weight parameter λ is tuned on the development set. To integrate the predict-
self reestimation into the iterative transformation training, a reversed transformation
model is introduced and the enhanced scoring function is used when the function
TRANSANNOTATE invokes the function DECODE.

5. Experiments

To evaluate the performance of annotation adaptation, we experiment on two impor-
tant NLP tasks, Chinese word segmentation and dependency parsing, both of which
can be modeled as discriminative classiﬁcation problems. For both tasks, we give the
performances of the baseline models and the annotation adaptation algorithms.

5.1 Experimental Set-ups

We perform annotation adaptation for word segmentation from People’s Daily (PD)
(Yu et al. 2001) to Penn Chinese Treebank 5.0 (CTB) (Xue et al. 2005). The two corpora
are built according to different segmentation guidelines and differ largely in quantity
of data. CTB is smaller in size with about 0.5M words, whereas PD is much larger,
containing nearly 6M words. Tableau 4 shows the data partitioning for the two corpora.

132

D
o
w
n
o
un
d
e
d

F
r
o
m
h

t
t

:
/
/

d
je
r
e
c
t
.

je
t
.

e
d
toi
/
c
o

je
je
/

un
r
t
je
c
e
–
p
d

F
/

4
1
1
1
1
9
1
8
0
5
3
5
3
/
c
o

je
je

_
un
_
0
0
2
1
0
p
d

b
oui
g
toi
e
s
t

o
n
0
8
S
e
p
e
m
b
e
r
2
0
2
3

Jiang et al.

Automatic Adaptation of Annotations

Tableau 4
Data partitioning for CTB and PD.

Partition

Sections

# of Words

CTB
Training

Developing
Test

PD
Training
Test

1 − 270
400 − 931
1, 001 − 1, 151
301 − 325
271 − 300

02 − 06
01

0.47M.

6.66K
7.82K

5.86M.
1.07M.

To approximate more general scenarios of annotation adaptation problems, nous
extract from PD a subset that is comparable to CTB in size. Because there are many
extremely long sentences in the original PD corpus, we ﬁrst split them into normal sen-
tences according to the full-stop punctuation symbol. We randomly select 20, 000 sen-
tences (0.45M words) from the PD training data as the new training set, et 1, 000/1, 000
sentences from the PD test data as the new testing/developing set. We label the smaller
version of PD as SPD. The balanced source corpus and target corpus also facilitate the
investigation of annotation adaptation.

Annotation adaptation for dependency parsing is performed from the CTB-derived
syntactic dependency treebank (DCTB) (Yamada and Matsumoto 2003) to the Seman-
tic Dependency Treebank (SDT) (Che et al. 2012). Semantic dependency encodes the
semantic relationships between words, which are very different from syntactic depen-
dencies. SDT is annotated on a small portion of the CTB text as depicted in Table 5;
donc, we use the subset of DCTB covering the remaining CTB text as the source
corpus. We still denote the source corpus as DCTB in the following for simplicity.

5.2 Baseline Segmenters and Parsers

We train the baseline perceptron classiﬁers for Chinese word segmentation on the train-
ing sets of CTB and SPD, using corresponding development sets to determine the best

Tableau 5
Data partitioning for SDT.

Partition

Sections

# of Words

Training

Developing

Test

1 − 10
36 − 65
81 − 121
1, 001 − 1, 078
1, 100 − 1, 119
1, 126 − 1, 140
66 − 80
1, 120 − 1, 125
11 − 35
1, 141 − 1, 151

244.44K

14.97K

33.51K

133

D
o
w
n
o
un
d
e
d

F
r
o
m
h

t
t

:
/
/

d
je
r
e
c
t
.

je
t
.

e
d
toi
/
c
o

je
je
/

un
r
t
je
c
e
–
p
d

F
/

4
1
1
1
1
9
1
8
0
5
3
5
3
/
c
o

je
je

_
un
_
0
0
2
1
0
p
d

b
oui
g
toi
e
s
t

o
n
0
8
S
e
p
e
m
b
e
r
2
0
2
3

Computational Linguistics

Volume 41, Nombre 1

Chiffre 6
Learning curve of the averaged perceptron classiﬁer on the CTB developing set.

training iterations. The performance measurement indicator for word segmentation is
the balanced F-measure, F = 2PR/(P. + R.), a function of Precision P and Recall R, où
P is the percentage of words in segmentation results that are segmented correctly, et
R is the percentage of correctly segmented words in the gold standard words.

For both syntactic and semantic dependency parsing, we concentrate on nonlabeled
parsing that predicts the graphic dependency structure for the input sentence without
considering dependency labels. The perceptron-based baseline dependency models are
trained on the training sets of DCTB and SDT, using the development sets to determine
the best training iterations. The performance measurement indicator for dependency
parsing is the Unlabeled Attachment Score, denoted as Precision P, indicating the
percentage of words in predicted dependency structure that are correctly attached to
their head words.

Chiffre 6 shows the learning curve of the averaged perceptron for word segmenta-
tion on the development set. Accuracies of the baseline classiﬁers are listed in Table 6.
We also report the performance of the classiﬁers on the testing sets of the opposite
corpora. Experimental results are in line with our expectations. A classiﬁer performs
better in its corresponding testing set, and performs signiﬁcantly worse on testing data
following a different annotation guideline.

Tableau 7 shows the accuracies of the baseline syntactic and semantic parsers, aussi
as the performance of the parsers on the testing sets of the opposite corpora. Similar

Tableau 6
Performance of the baseline word segmenters.

Test on (F1%)

Train on

CTB

SPD

97.35
91.23(↓ 3.02)

86.65(↓ 10.70)
94.25

CTB
SPD

134

D
o
w
n
o
un
d
e
d

F
r
o
m
h

t
t

:
/
/

d
je
r
e
c
t
.

je
t
.

e
d
toi
/
c
o

je
je
/

un
r
t
je
c
e
–
p
d

F
/

4
1
1
1
1
9
1
8
0
5
3
5
3
/
c
o

je
je

_
un
_
0
0
2
1
0
p
d

b
oui
g
toi
e
s
t

o
n
0
8
S
e
p
e
m
b
e
r
2
0
2
3

Jiang et al.

Automatic Adaptation of Annotations

Tableau 7
Performance of the baseline dependency parsers.

Test on (P%)

Train on

SDT

DCTB

SDT
DCTB

77.51
51.92(↓ 32.62)

53.62(↓ 23.89)
84.54

Tableau 8
Performance of automatic annotation adaptation models for word segmentation.

Model

Time (sec) Accuracy (F1%)

Merging

Model 1
Model 2
Model 3

Baseline

1.33

4.39
1.33
1.33

1.21

93.79

97.67
97.69
97.97

97.35

to the situations in word segmentation, two parsers give state-of-the-art accuracies on
their own testing sets, but perform poorly on the other testing sets. This indicates the
degree of divergence between the annotation guidelines of DCTB and SDT.

5.3 Automatic Annotation Adaptation

As a variant of Model 1, Model 2 shares the same transfer classiﬁer, and differs only in
training and decoding of the ﬁnal classiﬁer. Tables 8 et 9 show the performances of
systems resulting from Models 1 et 2, as well as the classiﬁers trained on the directly
merged corpora. The time costs for decoding are also listed to facilitate the practical
comparison.

We ﬁnd that the simple corpus merging strategy leads to a dramatic decrease in
accuracy, due to the different and incompatible annotation guidelines. Model 1, le
simplest model for annotation adaptation, gives signiﬁcant improvement over the
baseline classiﬁers for word segmentation and dependency parsing. This indicates that
the statistical regularities for annotation adaptation learned by the transfer classiﬁers

Tableau 9
Performance of automatic annotation adaptation models for dependency parsing.

Model

Time (min) Accuracy (P%)

Merging

Model 1
Model 2
Model 3

Baseline

7.38

15.43
7.38
7.38

6.67

65.95

78.69
78.73
79.34

77.51

135

D
o
w
n
o
un
d
e
d

F
r
o
m
h

t
t

:
/
/

d
je
r
e
c
t
.

je
t
.

e
d
toi
/
c
o

je
je
/

un
r
t
je
c
e
–
p
d

F
/

4
1
1
1
1
9
1
8
0
5
3
5
3
/
c
o

je
je

_
un
_
0
0
2
1
0
p
d

b
oui
g
toi
e
s
t

o
n
0
8
S
e
p
e
m
b
e
r
2
0
2
3

Computational Linguistics

Volume 41, Nombre 1

Chiffre 7
Training curve of the iterative training strategy over Model 2 for word segmentation.

bring performance improvement, utilizing guided decisions in the cascaded classiﬁers.
Model 2 leads to classiﬁers with accuracy increments comparable to those of Model 1,
while consuming only one third of the decoding time. It is inconsistent with our expec-
tation. The strategy of directly transforming the source corpus to the target guideline
also facilitates the utilization of more than one source corpus.

We ﬁrst introduce the iterative training strategy to Model 2. The corresponding
development sets are used to determine the best training iterations for the iterative
annotation transformations. After each iteration, we test the performance of the clas-
siﬁers trained on the merged corpora. Figures 7 et 8 show the performance curves
for Chinese word segmentation and semantic dependency parsing, respectivement, avec
iterations ranging from 1 à 10. The performance of Model 2 is naturally included

)

%
P.

(

oui
c
un
r
toi
c
c
UN

79.5

78.5

+ iterative training

Model 2 (baseline)

Training iterations

Chiffre 8
Training curve of the iterative training strategy over Model 2 for dependency parsing.

136

D
o
w
n
o
un
d
e
d

F
r
o
m
h

t
t

:
/
/

d
je
r
e
c
t
.

je
t
.

e
d
toi
/
c
o

je
je
/

un
r
t
je
c
e
–
p
d

F
/

4
1
1
1
1
9
1
8
0
5
3
5
3
/
c
o

je
je

_
un
_
0
0
2
1
0
p
d

b
oui
g
toi
e
s
t

o
n
0
8
S
e
p
e
m
b
e
r
2
0
2
3

Jiang et al.

Automatic Adaptation of Annotations

Chiffre 9
Performance of predict-self ﬁltering and predict-self re-estimation over Model 2 for word
segmentation.

in the curves (located at iteration 1). The curves show that, for both segmentation
and parsing, the accuracies of the classiﬁers trained on the merged corpora consis-
tently improve in the earlier iterations (par exemple., from iteration 2 to iteration 5 for word
segmentation).

Experiments for introduction of predict-self ﬁltering and predict-self re-estimation
are shown in Figures 9 et 10. The curves show the performances of the predict-
self re-estimation with a series of weight parameters, ranging from 0 à 1 with step
0.05. Note that in both ﬁgures, the points at λ = 0 show the performances of Model 2.
We ﬁnd that predict-self ﬁltering brings a slight improvement over the baseline for

79.5

78.5

)

%
P.

(

oui
c
un
r
toi
c
c
UN

Model 2 + predict-self reestimation

Model 2 + predict-self filtering

Model 2 (baseline)

0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

Predict-self ratio

Chiffre 10
Performance of predict-self ﬁltering and predict-self re-estimation over Model 2 for dependency
parsing.

137

D
o
w
n
o
un
d
e
d

F
r
o
m
h

t
t

:
/
/

d
je
r
e
c
t
.

je
t
.

e
d
toi
/
c
o

je
je
/

un
r
t
je
c
e
–
p
d

F
/

4
1
1
1
1
9
1
8
0
5
3
5
3
/
c
o

je
je

_
un
_
0
0
2
1
0
p
d

b
oui
g
toi
e
s
t

o
n
0
8
S
e
p
e
m
b
e
r
2
0
2
3

Computational Linguistics

Volume 41, Nombre 1

Chiffre 11
Training curve of annotation adaptation with Model 3 for word segmentation.

)

%
P.

(

oui
c
un
r
toi
c
c
UN

80.5

79.5

78.5

Model 3

+ iterative training

Model 2 (baseline)

Training iterations

Chiffre 12
Training curve of annotation adaptation with Model 3 for dependency parsing.

word segmentation, but even decreases the accuracy for dependency parsing. An initial
analysis on the experimental results reveals that the ﬁltering strategy discards some
complicated sentences in the source corpora, and the discarded sentences would bring
further improvement if properly used. Par exemple, in word segmentation, predict-self
ﬁltering discards 5% of sentences from the source corpus, containing nearly 10% of the
training words. For the two tasks, the predict-self re-estimation outperforms the ﬁltering
strategy. With properly tuned weights, predict-self re-estimation can make better use of
the training data. The largest accuracy improvements achieved over Model 2 for word
segmentation and dependency parsing are 0.3 points and 0.6 points.

Figures 11 et 12 show the performance curves after the introduction of both
iterative training and predict-self re-estimation on the basis of Model 2 (this enhanced

138

D
o
w
n
o
un
d
e
d

F
r
o
m
h

t
t

:
/
/

d
je
r
e
c
t
.

je
t
.

e
d
toi
/
c
o

je
je
/

un
r
t
je
c
e
–
p
d

F
/

4
1
1
1
1
9
1
8
0
5
3
5
3
/
c
o

je
je

_
un
_
0
0
2
1
0
p
d

b
oui
g
toi
e
s
t

o
n
0
8
S
e
p
e
m
b
e
r
2
0
2
3

Jiang et al.

Automatic Adaptation of Annotations

Tableau 10
The performance of the enhanced annotation adaptation for word segmentation, compared with
previous work.

Model

SPD → CTB
PD → CTB

Previous Work

(Jiang et al. 2008)
(Kruengkrai et al. 2009)
(Zhang and Clark 2010)
(Sun 2011)

Accuracy (F1%)

97.97
98.43

97.85
97.87
97.79
98.17

Tableau 11
The performance of the enhanced annotation adaptation for dependency parsing, compared
with the results of SemEval-2012 contest, where the best systems of each institute are listed.

Model

DCTB → SDT

SemEval-2012 Contest

Zhijun Wu-1
Zhou Qiaoli-3
NJU-Parser-1
ICT-1
Giuseppe Attardi-SVM-1-R

Accuracy (P%)

79.34

80.64
80.60
80.35
73.20
60.83

model is denoted as Model 3). We ﬁnd that the predict-self re-estimation brings im-
provement to the iterative training at each iteration, for both word segmentation and
dependency parsing. The maximum performance is achieved at iteration 4 for word
segmentation, and at iteration 5 for dependency parsing. The corresponding models
are evaluated on the corresponding testing sets, and the experimental results are also
shown in Tables 8 et 9. Compared to Model 1, the optimized annotation adaptation
strategy, Model 3, leads to classiﬁers with signiﬁcantly higher accuracy and to process-
ing speeds that are several times faster. Tables 10 et 11 show the experimental results
compared with previous work. For both Chinese word segmentation and semantic de-
pendency parsing, automatic annotation adaptation yields state-of-the-art performance,
despite using single classiﬁers with only local features. Note that for the systems in the
SemEval contest (Che et al. 2012), many other technologies including clause segmen-
tation, system combination, and complicated features were adopted, as well as elab-
orate engineering. We also performed signiﬁcance tests2 to verify the effectiveness of
annotation adaptation.We ﬁnd that for both Chinese word segmentation and semantic
dependency parsing, annotation adaptation brings signiﬁcant improvement (p < 0.001) over the baselines trained on the target corpora only. 2 http://www.cis.upenn.edu/∼dbikel/download/compare.pl. 139 l D o w n o a d e d f r o m h t t p : / / d i r e c t . m i t . e d u / c o l i / l a r t i c e - p d f / / / / 4 1 1 1 1 9 1 8 0 5 3 5 3 / c o l i _ a _ 0 0 2 1 0 p d . f b y g u e s t t o n 0 8 S e p e m b e r 2 0 2 3 Computational Linguistics Volume 41, Number 1 Table 12 Quantitative analysis for performance of annotation adaptation on word segmentation. The words are grouped according to POS tags, and for each group, the recall values of baseline and annotation adaptation are reported. To save space, only the categories with proportions of more than 1% and with performance ﬂuctuations of more than 0.1 points are listed. Word Type Proportion Baseline Anno. Ada. Trend AD CD JJ LC NN NR VA VV 3.99 2.07 3.25 1.36 29.95 9.81 1.58 13.52 98.75 98.79 95.01 98.16 97.79 97.32 95.27 98.33 96.87 99.39 95.40 99.08 98.83 98.98 97.63 98.61 ↓ ↑ ↑ ↑ ↑ ↑ ↑ ↑ To evaluate the stability of annotation adaptation, we perform quantitative analysis on the results of annotation adaptation. For word segmentation, the words are grouped according to POS tags. For dependency parsing, the dependency edges are grouped according to POS tag pairs. For each category, the recall values of baseline and annota- tion adaptation are reported. To ﬁlter the lists, we set two signiﬁcance thresholds with respect to the proportion of a category and the performance ﬂuctuation between two systems. For word segmentation, only the categories with proportions of more than 1% and with ﬂuctuations of more than 0.1 points are reserved, and for dependency parsing, the two thresholds are 1% and 0.5. Tables 12 and 13 show the analysis results for word segmentation and dependency parsing, respectively. For both tasks, annotation adaptation brings improvement for most of the situations. Table 13 Quantitative analysis for performance of annotation adaptation on dependency parsing. The dependency edges are grouped according to POS tag pairs, and for each group, the recall values of baseline and annotation adaptation are reported. To save space, only the categories with proportions of more than 1% and with performance ﬂuctuations of more than 0.5 points are listed. Edge Type Proportion Baseline Anno. Ada. Trend 2.61 1.00 1.68 3.14 2.31 3.15 1.89 1.33 1.07 5.55 1.76 11.36 1.71 9.89 7.72 90.76 98.19 83.48 79.82 85.49 75.71 64.64 67.87 77.80 81.69 93.82 78.85 76.19 64.40 59.74 93.41 97.29 87.43 83.86 84.83 76.39 71.81 70.13 82.86 83.00 94.51 82.02 79.01 66.91 62.21 ↑ ↓ ↑ ↑ ↓ ↑ ↑ ↑ ↑ ↑ ↑ ↑ ↑ ↑ ↑ NN→JJ NN→LC NN→M NN→NR NN→P NN→PU NN→VV VC→PU VE→NN VV→AD VV→DEC VV→NN VV→NR VV→PU VV→VV 140 l D o w n o a d e d f r o m h t t p : / / d i r e c t . m i t . e d u / c o l i / l a r t i c e - p d f / / / / 4 1 1 1 1 9 1 8 0 5 3 5 3 / c o l i _ a _ 0 0 2 1 0 p d . f b y g u e s t t o n 0 8 S e p e m b e r 2 0 2 3 Jiang et al. Automatic Adaptation of Annotations ) % F ( y c a r u c c A 98 97 96 95 94 93 92 91 90 Target corpus only Annotation Adaptation 1000 2000 4000 8000 16000 Scale of target corpus (# of sentences) Figure 13 Performance of annotation adaptation with varying-size target corpora for Chinese word segmentation. We further investigate the effect of varying the sizes of the target corpora. Exper- iments are conducted for word segmentation and dependency parsing with ﬁxed-size source corpora and varying-size target corpora. We use SPD and DCTB as the source corpora for word segmentation and dependency parsing, respectively. Figures 13 and 14 show the performance curves on the testing sets. We ﬁnd that, for both word segmen- tation and dependency parsing, the improvements brought by annotation adaptation are more signiﬁcant when the target corpora are smaller. It means that the automatic annotation adaptation is more valuable when the size of the target corpus is small, ) % P ( y c a r u c c A 80 79 78 77 76 75 74 73 72 71 70 Target corpus only Annotation Adaptation 500 1000 2000 4000 8000 Scale of target corpus (# of sentences) Figure 14 Performance of annotation adaptation with varying-size target corpora for semantic dependency parsing. 141 l D o w n o a d e d f r o m h t t p : / / d i r e c t . m i t . e d u / c o l i / l a r t i c e - p d f / / / / 4 1 1 1 1 9 1 8 0 5 3 5 3 / c o l i _ a _ 0 0 2 1 0 p d . f b y g u e s t t o n 0 8 S e p e m b e r 2 0 2 3 Computational Linguistics Volume 41, Number 1 which is good news for the situation where the corpus we are concerned with is smaller but a larger differently annotated corpus exists. Of course, the comparison between automatic annotation adaptation and previous strategies without using additional training data is unfair. Our work aims to ﬁnd another way to improve NLP tasks: focusing on the collection of more training data instead of making full use of a certain corpus. We believe that the performance of automatic annotation adaptation can be further improved by adopting the advanced technologies of previous work, such as complicated features and model combination. It would be useful to conduct experiments with more source-annotated training data, such as the SIGHAN data set for word segmentation, to investigate the trend of improvement along with the further increment of annotated sentences. It would also be valuable to evaluate the improved word segmenter and dependency parser on the out-of-domain data sets. However, currently most corpora for word segmentation and dependency parsing do not explicitly distinguish the domains of their data sections, making such evaluations difﬁcult to conduct. 6. Discussion: Application Situations Automatic annotation adaptation aims to transform the annotations in a corpus to the annotations following other guidelines. The models for annotation adaptation use a transfer classiﬁer to learn the statistical correspondence regularities between different annotation guidelines. These statistical regularities are learned from a parallel anno- tated corpus, which does not need to be manually annotated. In fact, the models for annotation adaptation train the transfer classiﬁer on an automatically generated parallel annotated corpus, which is generated by processing a corpus with a classiﬁer trained on another corpus. That is to say, if we want to conduct annotation adaptation across several corpora, no additional corpora need to be manually annotated. This setting makes the strategy of annotation adaptation more general, because it is much harder to manually annotate a parallel annotated corpus, regardless of the language or the NLP problem under consideration. To tackle the problem of noise in automatically generated annotations, the advanced models we designed generate a better parallel annotated corpus by making use of strategies such as iterative optimization. Automatic annotation adaptation can be applied in any situation where we have multiple corpora with different and incompatible annotation philosophies for the same task. As our case studies, both Chinese word segmentation and dependency parsing have more than one corpora with different annotation guidelines, such as the People’s Daily and the Penn Chinese Treebank for Chinese word segmentation. In a more abstract view, constituency grammar and dependency grammar can be treated as two annota- tion guidelines for parsing. The syntactic knowledge in a constituency treebank and a dependency treebank, therefore, can be integrated by automatic annotation adaptation. For example, the LinGo Redwoods Treebank can also be transformed to the annotation guideline of the Semantic Dependency Treebank. Furthermore, the annotations (such as a grammar) given by bilingual projection or unsupervised induction can be seen as following a special annotation philosophy. For bilingually projected annotations, the annotation guideline would be similar to that of the counterpart language. For unsupervised induced annotations, the annotation guide- line reﬂects the statistical structural distribution of a speciﬁc data set. In both situations, the underlying annotation guidelines may be largely different from that of the testing sets, which usually come from human-annotated corpora. The system trained on a bilin- gually projected or unsupervised induced corpus may perform poorly on an existing 142 l D o w n o a d e d f r o m h t t p : / / d i r e c t . m i t . e d u / c o l i / l a r t i c e - p d f / / / / 4 1 1 1 1 9 1 8 0 5 3 5 3 / c o l i _ a _ 0 0 2 1 0 p d . f b y g u e s t t o n 0 8 S e p e m b e r 2 0 2 3 Jiang et al. Automatic Adaptation of Annotations testing set, but if the projected or induced corpus has high inner consistency, it could improve a system trained on an existing corpus by automatic annotation adaptation. In this point of view, the practical value of the current work on bilingual projection and unsupervised induction may be underestimated, and annotation adaptation could make better use of the projected or induced knowledge.3 7. Related Work There has already been some preliminary work tackling the divergence between dif- ferent annotation guidelines. Gao et al. (2004) described a transformation-based con- verter to transfer a certain word segmentation result to another annotation guideline. They designed class-type transformation templates and used the transformation-based error-driven learning method of Brill (1995) to learn what word delimiters should be modiﬁed. Many efforts have been devoted to manual treebank transformation, where PTB is adapted to other grammar formalisms, such as CCG and LFG (Cahill et al. 2002; Hockenmaier and Steedman 2007). However, all these are heuristic-based—that is, they need manually designed transformation templates and involve heavy human engineering. Such strategies are hard to be generalized to POS tagging, not to mention other complicated structural prediction tasks. We investigated the automatic integration of word segmentation knowledge in differently annotated corpora (Jiang, Huang, and Liu 2009; Jiang et al. 2012), which can be seen as the preliminary work of automatic annotation adaptation. Motivated by our initial investigation, researchers applied similar methodologies to constituency parsing (Sun, Wang, and Zhang 2010; Zhu, Zhu, and Hu 2011) and word segmentation (Sun and Wan 2012). This previous work veriﬁed the effectiveness of automatic annotation adaptation, but did not reveal the essential deﬁnition of the problem nor the intrinsic principles of the solutions. Instead, this work clearly deﬁnes the problem of annotation adaptation, reveals the intrinsic principles of the solutions, and systematically describes a series of gradually improved models. The most advanced model learns transformation regularities much better and achieves signiﬁcant higher accuracy for both word segmen- tation and dependency parsing, without slowing down the ﬁnal language processors. The problem of automatic annotation adaptation can be seen as a special case of transfer learning (Pan and Yang 2010), where the source and target tasks are similar, but not identical. More speciﬁcally, the problem related to annotation adaptation assumes that the labeling mechanism across the source and target tasks are the same, but the predictive functions are different. The goal of annotation adaptation is to adapt the source predictive function to be used for the target task by exploiting the labeled data of the source task and the target task. Furthermore, automatic annotation adaptation approximately falls into the spectrum of relational-knowledge-transfer problems (Mihalkova, Huynh, and Mooney 2007; Mihalkova and Mooney 2008; Davis and Domingos 2009), but it tackles problems where the relations among data between the source and target domains can be largely isomerous—or, in other words, with different and incompatible annotation schemes. This work enriches the research of transfer learning by proposing and solving an NLP problem different from previous situations. For more details of transfer learning please refer to the survey of Pan and Yang (2010). 3 We have performed preliminary experiments on word segmentation. Bilingual projection was conducted from English to Chinese with the Chinese–English FBIS as the bilingual corpus. By annotation adaptation, the projected corpus for word segmentation brings a signiﬁcant F-measure increment of nearly 0.6 points over the baseline trained on CTB only. 143 l D o w n o a d e d f r o m h t t p : / / d i r e c t . m i t . e d u / c o l i / l a r t i c e - p d f / / / / 4 1 1 1 1 9 1 8 0 5 3 5 3 / c o l i _ a _ 0 0 2 1 0 p d . f b y g u e s t t o n 0 8 S e p e m b e r 2 0 2 3 Computational Linguistics Volume 41, Number 1 The training procedure for an annotation adaptation model requires a parallel an- notated corpus (which may be automatically generated); this fact puts the method into the neighborhood of the family of approaches known as annotation projection (Hwa et al. 2002, 2005; Ganchev, Gillenwater, and Taskar 2009; Smith and Eisner 2009; Jiang and Liu 2010; Das and Petrov 2011). Essentially, annotation adaptation and annotation projection tackle different problems; the former aims to transform the annotations from one guideline to another (of course in the same language), whereas the latter aims to project the annotation (as well as the annotation guideline) from one language to another. Therefore, the machine learning methods for annotation adaptation pay attention to automatic transformation of annotations, while for annotation projection, the machine learning methods focus on the bilingual projection across languages. Co-training (Sarkar 2001) and classiﬁer combination (Nivre and McDonald 2008) are two techniques for training improved dependency parsers. The co-training technique lets two different parsing models learn from each other during the parsing of unlabeled text: One model selects some unlabeled sentences it can conﬁdently parse, and provides them to the other model as additional training data in order to train more powerful parsers. The classiﬁer combination lets graph-based and transition-based dependency parsers utilize the features extracted from each other’s parsing results, to obtain com- bined, enhanced parsers. The two techniques aim to let two models learn from each other on the same corpus with the same distribution and annotation guideline, whereas our strategy aims to integrate the knowledge in multiple corpora with different annota- tion guidelines. The iterative training procedure used in the optimized model shares some sim- ilarities with the co-training algorithm in parsing (Sarkar 2001), where the training procedure lets two different models learn from each other during parsing of the raw text. The key idea of co-training is to utilize the complementarity of different parsing models to mine additional training data from raw text, whereas iterative training for annotation adaptation emphasizes the iterative optimization of the parallel annotated corpora used to train the transfer classiﬁers. The predict-self methodology is implicit in many unsupervised learning approaches; it has been successfully used in unsupervised dependency parsing (Daum´e III 2009). We adapt this idea to the scenario of annotation adaptation to improve transformation accuracy. In recent years much effort has been devoted to the improvement of word seg- mentation and dependency parsing. For example, the introduction of global training or complicated features (Zhang and Clark 2007, 2010); the investigation of word struc- tures (Li 2011); the strategies of hybrid, joint, or stacked modeling (Nakagawa and Uchimoto 2007; Kruengkrai et al. 2009; Wang, Zong, and Su 2010; Sun 2011); and the semi-supervised and unsupervised technologies utilizing raw text (Zhao and Kit 2008; Johnson and Goldwater 2009; Mochihashi, Yamada, and Ueda 2009; Hewlett and Cohen 2011). We believe that the annotation adaptation technologies can be adopted jointly with complicated features, system combination, and semi-supervised/unsupervised technologies to further improve the performance of word segmentation and depen- dency parsing. 8. Conclusion and Future Work We have described the problem of annotation adaptation and the intrinsic principles of its solutions, and proposed a series of successively enhanced models that can au- tomatically adapt the divergence between different annotation formats. These models learn the statistical regularities of adaptation between different annotation guidelines, 144 l D o w n o a d e d f r o m h t t p : / / d i r e c t . m i t . e d u / c o l i / l a r t i c e - p d f / / / / 4 1 1 1 1 9 1 8 0 5 3 5 3 / c o l i _ a _ 0 0 2 1 0 p d . f b y g u e s t t o n 0 8 S e p e m b e r 2 0 2 3 Jiang et al. Automatic Adaptation of Annotations and integrate the knowledge in corpora with different annotation guidelines. In the problems of Chinese word segmentation and semantic dependency parsing, annotation adaptation algorithms bring signiﬁcant improvements by integrating the knowledge in differently annotated corpora, People’s Daily and Penn Chinese Treebank for word seg- mentation, and Penn Chinese Treebank and Semantic Dependency Parsing for depen- dency parsing. For both tasks, annotation adaptation leads to a segmenter and a parser achieving the state-of-the-art, despite using only local features in single classiﬁers. Many aspects related to annotation adaptation deserve further investigation in the future. First, models for annotation adaptation can be adapted to other NLP tasks such as semantic analysis. Second, jointly tackling the divergences in both annotations and domains is an important problem. In addition, an unsupervised-induced or bilingually projected corpus, despite performing poorly on the speciﬁed testing data, may have high inner annotation consistency. That is to say, the induced corpora can be treated as a knowledge source following another annotation guideline, and the performance of current unsupervised or bilingually projected models may be seriously underestimated. Annotation adaptation may give us a new perspective on knowledge induction and measurement for such methods. Acknowledgments Jiang, L ¨u, and Liu were supported by National Natural Science Foundation of China (contract 61202216) and the National Key Technology R&D Program (no. 2012BAH39B03). Huang was supported in part by the DARPA DEFT Project (FA8750-13-2-0041). Liu was partially supported by the Science Foundation Ireland (grant no. 07/CE/I1142) as part of the CNGL at Dublin City University. We also thank the anonymous reviewers for their insightful comments. Finally, we want to thank Chris Hokamp for proofreading. References Blitzer, John, Ryan McDonald, and Fernando Pereira. 2006. Domain adaptation with structural correspondence learning. In Proceedings of EMNLP, pages 120–128, Sydney. Brill, Eric. 1995. Transformation-based error-driven learning and natural language processing: A case study in part-of-speech tagging. Computational Linguistics, 21(4):543–565. Buchholz, Sabine and Erwin Marsi. 2006. CONLL-X shared task on multilingual dependency parsing. In Proceedings of CoNLL, pages 149–164, New York, NY. Cahill, Aoife, Mairead McCarthy, Josef van Genabith, and Andy Way. 2002. Automatic annotation of the Penn treebank with LFG F-structure information. In Proceedings of the LREC Workshop, Las Palmas. Che, Wanxiang, Meishan Zhang, Yanqiu Shao, and Ting Liu. 2012. Semeval-2012 task 5: Chinese semantic dependency parsing. In Proceedings of SemEval, pages 378–384, Montreal. Collins, Michael. 2002. Discriminative training methods for hidden Markov models: Theory and experiments with perceptron algorithms. In Proceedings of EMNLP, pages 1–8, Philadelphia, PA. Das, Dipanjan and Slav Petrov. 2011. Unsupervised part-of-speech tagging with bilingual graph-based projections. In Proceedings of ACL, pages 600–609, Portland, OR. Daum´e III, Hal. 2007. Frustratingly easy domain adaptation. In Proceedings of ACL, pages 256–263, Prague. Daum´e III, Hal. 2009. Unsupervised search- based structured prediction. In Proceedings of ICML, pages 209–216, Montreal. Daum´e III, Hal and Daniel Marcu. 2006. Domain adaptation for statistical classiﬁers. Journal of Artiﬁcial Intelligence Research, 26:101–126. Davis, Jesse and Pedro Domingos. 2009. Deep transfer via second-order Markov logic. In Proceedings of ICML, pages 217–224, Montreal. Eisner, Jason M. 1996. Three new probabilistic models for dependency parsing: An exploration. In Proceedings of COLING, pages 340–345, Copenhagen. Ganchev, Kuzman, Jennifer Gillenwater, and Ben Taskar. 2009. Dependency grammar induction via bitext projection constraints. In Proceedings of ACL, pages 369–377, Singapore. 145 l D o w n o a d e d f r o m h t t p : / / d i r e c t . m i t . e d u / c o l i / l a r t i c e - p d f / / / / 4 1 1 1 1 9 1 8 0 5 3 5 3 / c o l i _ a _ 0 0 2 1 0 p d . f b y g u e s t t o n 0 8 S e p e m b e r 2 0 2 3 Computational Linguistics Volume 41, Number 1 Gao, Jianfeng, Andi Wu, Mu Li, Chang-Ning Huang, Hongqiao Li, Xinsong Xia, and Haowei Qin. 2004. Adaptive Chinese word segmentation. In Proceedings of ACL, pages 462–469, Barcelona. Hewlett, Daniel and Paul Cohen. 2011. Fully unsupervised word segmentation with BVE and MDL. In Proceedings of ACL, pages 540–545, Portland, OR. Hockenmaier, Julia and Mark Steedman. 2007. CCGBank: A corpus of CCG derivations and dependency structures extracted from the Penn treebank. Computational Linguistics, 33(3):355–396. Hwa, Rebecca, Philip Resnik, Amy Weinberg, Clara Cabezas, and Okan Kolak. 2005. Bootstrapping parsers via syntactic projection across parallel texts. Natural Language Engineering, 11(3):311–325. Hwa, Rebecca, Philip Resnik, Amy Weinberg, and Okan Kolak. 2002. Evaluating translational correspondence using annotation projection. In Proceedings of ACL, pages 392–399, Philadephia, PA. Jiang, Wenbin, Liang Huang, and Qun Liu. 2009. Automatic adaptation of annotation standards: Chinese word segmentation and POS tagging – A case study. In Proceedings of ACL, pages 522–530, Singapore. Jiang, Wenbin, Liang Huang, Yajuan L ¨u, and Qun Liu. 2008. A cascaded linear model for joint Chinese word segmentation and part-of-speech tagging. In Proceedings of ACL, pages 897–904, Columbus, OH. Jiang, Wenbin and Qun Liu. 2010. Dependency parsing and projection based on word-pair classiﬁcation. In Proceedings of the ACL, pages 12–20, Uppsala. Jiang, Wenbin, Fandong Meng, Qun Liu, and Yajuan L ¨u. 2012. Iterative annotation transformation with predict-self reestimation for Chinese word segmentation. In Proceedings of EMNLP, pages 412–420, Jeju Island. Johnson, Mark and Sharon Goldwater. 2009. Improving nonparameteric Bayesian inference: Experiments on unsupervised word segmentation with adaptor grammars. In Proceedings of NAACL, pages 317–325, Boulder, CO. Kruengkrai, Canasai, Kiyotaka Uchimoto, Junichi Kazama, Yiou Wang, Kentaro Torisawa, and Hitoshi Isahara. 2009. An error-driven word-character hybrid model for joint Chinese word segmentation and POS tagging. In Proceedings of ACL-IJCNLP, pages 513–521, Singapore. 146 Li, Zhongguo. 2011. Parsing the internal structure of words: A new paradigm for Chinese word segmentation. In Proceedings of ACL, pages 1,405–1,414, Portland, OR. Marcus, Mitchell P., Beatrice Santorini, and Mary Ann Marcinkiewicz. 1993. Building a large annotated corpus of English: The Penn treebank. Computational Linguistics, 19(2):313–330. Martins, Andr´e F. T., Dipanjan Das, Noah A. Smith, and Eric P. Xing. 2008. Stacking dependency parsers. In Proceedings of EMNLP, pages 157–166, Honolulu, HI. McDonald, Ryan, Koby Crammer, and Fernando Pereira. 2005. Online large-margin training of dependency parsers. In Proceedings of ACL, pages 91–98, Ann Arbor, MI. McDonald, Ryan and Fernando Pereira. 2006. Online learning of approximate dependency parsing algorithms. In Proceedings of EACL, pages 81–88, Trento. Mihalkova, Lilyana, Tuyen Huynh, and Raymond J. Mooney. 2007. Mapping and revising Markov logic networks for transfer learning. In Proceedings of AAAI, volume 7, pages 608–614, Vancouver. Mihalkova, Lilyana and Raymond J. Mooney. 2008. Transfer learning by mapping with minimal target data. In Proceedings of AAAI Workshop Transfer Learning for Complex Tasks, Chicago, IL. Mochihashi, Daichi, Takeshi Yamada, and Naonori Ueda. 2009. Bayesian unsupervised word segmentation with nested Pitman-Yor language modeling. In Proceedings of ACL-IJCNLP, pages 100–108, Singapore. Nakagawa, Tetsuji and Kiyotaka Uchimoto. 2007. A hybrid approach to word segmentation and POS tagging. In Proceedings of ACL, pages 217–220, Prague. Ng, Hwee Tou and Jin Kiat Low. 2004. Chinese part-of-speech tagging: One-at-a-time or all-at-once? Word-based or character-based? In Proceedings of EMNLP, pages 277–284, Barcelona. Nivre, Joakim and Ryan McDonald. 2008. Integrating graph-based and transition-based dependency parsers. In Proceedings of ACL, pages 950–958, Columbus, OH. Oepen, Stephan, Kristina Toutanova, Stuart Shieber, Christopher Manning Dan Flickinger, and Thorsten Brants. 2002. The LinGo Redwoods treebank: Motivation and preliminary applications. In Proceedings of COLING, volume 2, pages 1–5, Taipei. l D o w n o a d e d f r o m h t t p : / / d i r e c t . m i t . e d u / c o l i / l a r t i c e - p d f / / / / 4 1 1 1 1 9 1 8 0 5 3 5 3 / c o l i _ a _ 0 0 2 1 0 p d . f b y g u e s t t o n 0 8 S e p e m b e r 2 0 2 3 Jiang et al. Automatic Adaptation of Annotations Pan, Sinno Jialin and Qiang Yang. 2010. A survey on transfer learning. IEEE TKDE, 22(10):1345–1359. Sarkar, Anoop. 2001. Applying co-training methods to statistical parsing. In Proceedings of NAACL, pages 1–8, Pittsburgh, PA. Smith, David and Jason Eisner. 2009. Parser adaptation and projection with quasi-synchronous grammar features. In Proceedings of EMNLP, volume 2, pages 822–831, Singapore. Sun, Weiwei. 2011. A stacked sub-word model for joint Chinese word segmentation and part-of-speech tagging. In Proceedings of ACL, pages 1,385–1,394, Portland, OR. Sun, Weiwei and Xiaojun Wan. 2012. Reducing approximation and estimation errors for Chinese lexical processing with heterogeneous annotations. In Proceedings of ACL, volume 1, pages 232–241, Jeju Island. Sun, Weiwei, Rui Wang, and Yi Zhang. 2010. Discriminative parse reranking for Chinese with homogeneous and heterogeneous annotations. In Proceedings of CIPS-SIGHAN, Beijing. Available at http://aclweb.org/anthology/W10-4144. Wang, Kun, Chengqing Zong, and Keh-Yih Su. 2010. A character-based joint model for Chinese word segmentation. In Proceedings of COLING, pages 1,173–1,181, Beijing. Xue, Nianwen and Libin Shen. 2003. Chinese word segmentation as LMR tagging. In Proceedings of SIGHAN Workshop, volume 17, pages 176–179, Sapporo. Xue, Nianwen, Fei Xia, Fu-Dong Chiou, and Martha Palmer. 2005. The Penn Chinese treebank: Phrase structure annotation of a large corpus. Natural Language Engineering, 11(2):207–238. Yamada, H. and Y. Matsumoto. 2003. Statistical dependency analysis with support vector machines. In Proceedings of IWPT, pages 195–206, Nancy. Yu, Shiwen, Jianming Lu, Xuefeng Zhu, Huiming Duan, Shiyong Kang, Honglin Sun, Hui Wang, Qiang Zhao, and Weidong Zhan. 2001. Processing norms of modern Chinese corpus. Technical report, Institute of Computational Linguistics, Peking University. Zhang, Yue and Stephen Clark. 2007. Chinese segmentation with a word-based perceptron algorithm. In Proceedings of ACL, pages 840–847, Prague. Zhang, Yue and Stephen Clark. 2010. A fast decoder for joint word segmentation and POS-tagging using a single discriminative model. In Proceedings of EMNLP, pages 843–852, Cambridge, MA. Zhao, Hai and Chunyu Kit. 2008. Unsupervised segmentation helps supervised learning of character tagging for word segmentation and named entity recognition. In Proceedings of IJCNLP, pages 106–111, Hyderabad. Zhu, Muhua, Jingbo Zhu, and Minghan Hu. 2011. Better automatic treebank conversion using a feature-based approach. In Proceedings of ACL, volume 2, pages 715–719, Portland, OR. 147 l D o w n o a d e d f r o m h t t p : / / d i r e c t . m i t . e d u / c o l i / l a r t i c e - p d f / / / / 4 1 1 1 1 9 1 8 0 5 3 5 3 / c o l i _ a _ 0 0 2 1 0 p d . f b y g u e s t t o n 0 8 S e p e m b e r 2 0 2 3 l D o w n o a d e d f r o m h t t p : / / d i r e c t . m i t . e d u / c o l i / l a r t i c e - p d f / / / / 4 1 1 1 1 9 1 8 0 5 3 5 3 / c o l i _ a _ 0 0 2 1 0 p d . f b y g u e s t t o n 0 8 S e p e m b e r 2 0 2 3
Télécharger le PDF