Word Reordering and Comma Insertion Integrated with
Shift-Reduce Dependency Parsing
Kota Miyachi
1
, Tomohiro Ohno
2
a
and Shigeki Matsubara
3 b
1
Graduate School of Informatics, Nagoya University, Nagoya, Japan
2
Graduate School of Advanced Science and Technology, Tokyo Denki University, Tokyo, Japan
3
Information & Communications, Nagoya University, Nagoya, Japan
Keywords:
Sentence Generation, Parsing, Paraphrasing, Readability, Punctuation Insertion.
Abstract:
Japanese has widely recognized as relatively free word order language. However, since Japanese word order
is not completely arbitrary and has some sort of preference, even native Japanese writers often write Japanese
sentences that are grammatically well-formed but not easy to read. Furthermore, in Japanese sentences, a
comma plays an important role in explicitly separating the constituents such as words and phrases. This
paper proposes a method of word reordering and comma insertion for hard-to-read Japanese sentences so that
they become easier to read, as basic technique. Our contribution is to show the feasibility of concurrently
performing word reordering, comma insertion, and dependency parsing.
1 INTRODUCTION
Japanese has a relatively free word order structure.
This characteristic makes the possibility of writing
meaningful sentences without paying much attention
to word order. In practice, however, the existence of
word order preferences can lead to grammatically cor-
rect but hard-to-read sentences. Similarly, commas
are relatively freely inserted into a Japanese sentence,
but there are preferences, so it is necessary to place
commas in the right places to make a sentence easy to
read.
Reordering of words, which arranges a sentence
into readable word order, has been studied as a ba-
sic technique for several applications such as writ-
ing assistance and sentence generation. Yoshida et al.
(2014) proposed a word reordering method by con-
currently executing it with dependency parsing. How-
ever, in this method, commas are not considered.
On the other hand, Murata et al. (2010) developed a
Japanese comma insertion method by analyzing fea-
tures which capture the tendency of commas appear-
ing in Japanese texts. Since this method uses depen-
dency information, if the word order of an input sen-
tence is hard-to-read, the accuracy of comma inser-
tion would decrease. The reason is that an increase of
a
https://orcid.org/0000-0001-7015-7714
b
https://orcid.org/0000-0003-0416-3635
errors of dependency parsing affect the accuracy.
In this paper, we propose a method of simultane-
ous execution of dependency parsing, word reorder-
ing and comma insertion for hard-to-read Japanese
sentences. With their simultaneous execution, we can
analyze the dependencies and comma positions of an
input sentence considering not only the word order of
the input sentence but also the sequence of words after
reordering.
This paper is organized as follows: Section 2 ex-
plains aspects of word order, comma, and dependency
in Japanese. In Section 3, we propose a method
of word reordering and comma insertion integrated
with shift-reduce dependency parsing, while Section
4 conducts an experiment to evaluate the proposed
method quantitatively. Finally, Section 5 summarizes
the paper.
2 WORD ORDER, COMMA,
DEPENDENCY IN JAPANESE
The word order in a sentence relates to dependencies
of the sentence. Uchimoto et al. (2000) have con-
ducted word reordering using the syntactic informa-
tion based on the premise that dependency parsing
has been preliminarily performed, as well as (Belz
and Kow, 2011; Filippova and Strube, 2007; Har-
1144
Miyachi, K., Ohno, T. and Matsubara, S.
Word Reordering and Comma Insertion Integrated with Shift-Reduce Dependency Parsing.
DOI: 10.5220/0010335911441150
In Proceedings of the 13th International Conference on Agents and Artificial Intelligence (ICAART 2021) - Volume 2, pages 1144-1150
ISBN: 978-989-758-484-8
Copyright
c
2021 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved
7KHVHQWHQFHVDDQGEKDYHWKHVDPHPHDQLQJWKDWLVWUDQVODWHGDVn,OHIWKRPHORQJLQJIRUWKHFLW\|LQ(QJOLVK
7KHGLIIHUHQFHEHWZHHQDDQGELVMXVWLQWKHLUZRUGRUGHUVDQGFRPPDSRVLWLRQVLQ-DSDQHVH
ᶅ
ࢴͺ
,
ᶈ
ಎΗ
ORQJLQJIRU
ᶆ
ՊΝ
KRPH
ᶇ
ճͶ
WKHFLW\
ᶉ
ड़ͪɽ
OHIW
ɼ
ERXWSXWDVHQWHQFHWKDWKDVDSSURSULDWHZRUGRUGHUDQGFRPPDSRVLWLRQVDQGLWVGHSHQGHQF\VWUXFWXUH
ᶅ
ࢴͺ
,
ᶈ
ಎΗ
ORQJLQJIRU
ᶆ
ՊΝ
KRPH
ᶇ
ճͶ
WKHFLW\
ᶉ
ड़ͪɽ
OHIW
DLQSXWDVHQWHQFHWKDWKDVLQDSSURSULDWHZRUGRUGHUDQGFRPPDSRVLWLRQV
Figure 1: An example of the input and its ideal output in our method.
busch et al., 2006; Kruijff et al., 2001; Ringger et al.,
2004; Shaw and Hatzivassiloglou, 1999; Yokobayashi
et al., 2004). However, since dependency parsers
are usually trained on syntactically annotated corpora
in which sentences have the appropriate word order
(e.g., newspaper articles sentences); the accuracy of
dependency parsing is reduced when the input sen-
tences are not easy to read because of their word or-
ders. For the same reason, comma insertion is also
reduced.
On the other hand, if the word reordering is con-
ducted before the dependency parsing, the accuracy
of the word order would be decrease because the de-
pendency information cannot be used. In addition,
since comma information is used in dependency pars-
ing, the accuracy of the dependency parsing might be
decreased if dependency parsing is performed before
comma insertion.
As mentioned above, it can be concluded that de-
pendencyparsing, word reordering, and comma inser-
tion are interdependent. Therefore, as an approach to
the tasks of word reordering and comma insertion for
hard-to-read sentences due to the word order, it seems
suitable to perform them concurrently.
3 PROPOSED METHOD
In our method, we assume that the input sentence
might have inappropriate word order and comma po-
sitions, which causes not easy to read it but gram-
matically well-formed (e.g., Figure 1(a)). Based on
the assumption, our method simultaneously identifies
the dependency structure, appropriate word order and
comma positions in the sentence (e.g., Figure 1(b)).
As one of the strategies for executing these three
processes simultaneously, we can think of extending
(Yoshida et al., 2014), which achieved simultaneous
execution of dependency parsing and word reorder-
ing, by concurrently performing also comma inser-
tion. In other words, the strategy is to search for
the most likely pattern among all possible patterns of
dependency structure, word order, and comma posi-
tions for a whole sentence using dynamic program-
ming. However, adopting such a strategy needs to
consider all possible patterns, including comma posi-
tions, along with the dependency structure and word
order. Due to this fact, the potential number of such
patterns is too huge. Basides, it is not easy to search
for maximum likelihood patterns efficiently because
dynamic programming cannot be simply applied due
to the non-independent nature of each process.
Our method realizes the simultaneous execution
of dependency parsing, word reordering, and comma
insertion by focusing on two local bunsetsus
1
in the
input Japanese sentence, and determining the depen-
dency, word order, and existence or non-existence of
a comma between the two local bunsetsus.
3.1 Algorithm
In this study, we extend the Shift-Reduce algorithm
for Japanese dependency parsing (Sassano, 2004) to
realize the strategy of simultaneously determining the
dependency, word order, and comma between two lo-
cal bunsetsus in an input sentence.
The Shift-Reduce algorithm (Sassano, 2004) de-
termines the dependency structure of a whole sen-
tence by repeating the choice of two operators, Shift
and Reduce, depending on whether the bunsetsu in
the top of the stack depends on the bunsetsu in the
front of the queue or not. In this study, we extended
the algorithm by adding new operators for comma in-
1
Bunsetsu is a linguistic unit in Japanese that roughly
corresponds to a basic phrase in English. A bunsetsu con-
sists of one independent word and zero or more ancillary
words. A dependency relation in Japanese is a modification
relation in which a modifier bunsetsu depends on a modified
bunsetsu. That is, the modifier bunsetsu and the modified
bunsetsu work as modifier and modifyee, respectively
Word Reordering and Comma Insertion Integrated with Shift-Reduce Dependency Parsing
1145
6KLIW6WDFN
4XHXH
2SHUDWLRQ
6KLIW
6KLIW
6KLIW
5HGXFH
6ZDS 6WDFN
6ZDS
6&
5HGXFH
6KLIW
6KLIW
5HGXFH
5HGXFH
6KLIW
)LQ
5HGXFH
7LPH
%RWWRP %RWWRP7RS 7RS )URQW 5HDU
ᶅ
ᶈ
ᶆ
ᶇ
ᶉ
ʴ,OHIWKRPHORQJLQJIRUWKHFLW\ʵ
ᶅ
ᶈ
ᶆ
ᶇ
ᶉ
ᶅ
ᶈ
ᶆ
ᶇ
ᶉ
ᶅ
ᶈ
ᶆ
ᶉ
ᶇ
ᶉ
ᶅ
ᶇ ᶈ
ᶆ
ᶉ
ᶅ
ᶇ ᶈ
ᶆ
ᶉ
ᶅ
ᶇ
ᶈ
ᶆ
ᶉ
ᶅ
ᶇ
ᶈ
ᶆ
ᶉ
ᶅ
ᶇ
ᶈ
ᶆ
ᶉ
ᶅ
ᶇ
ᶈ
ᶆ
ᶉ
ᶅ
ᶇ
ᶈ
ᶆ
ᶉ
ᶅ
ᶇ
ᶈ
ᶆ
ᶉ
ᶅ
ᶇ
ᶈ
ᶅ
ࢴͺ
,
ᶈ
ಎΗ
ORQJLQJ
IRU
ᶆ
ՊΝ
KRPH
ᶇ
ճͶ
WKH
FLW\
ᶉ
ड़ͪɽ
OHIW
7KHRXWSXWWUHHH[SUHVVHVWKHIROORZLQJ
ZRUGRUGHUFRPPDVDQGGHSHQGHQF\VWUXFWXUH
ᶆ
ᶅ
ࢴͺ
,
ᶈ
ಎΗ
ORQJLQJ
IRU
ᶆ
ՊΝ
KRPH
ᶇ
ճͶ
WKH
FLW\
ᶉ
ड़ͪɽ
OHIW
ɼ
ɼ
ɼ
ɼ
ɼ
ɼ
ɼ
ɼ
ɼ
Figure 2: An example of simultaneous execution process of dependency parsing, word reordering, and comma insertion.
sertion (Shift-Comma and Reduce-Comma) and word
reordering (Swap) and a new stack for word reorder-
ing (swap stack), thus realized simultaneous execu-
tion of dependency parsing, word reordering, and
comma insertion. In the following, the stack orig-
inally used in the Shift-Reduce algorithm (Sassano,
2004) is called simply “shift stack” to distinguish it
from a new stack for word reordering (swap stack).
Figure 2 shows the simultaneous execution of de-
pendencyparsing, word reordering, and comma inser-
ICAART 2021 - 13th International Conference on Agents and Artificial Intelligence
1146
tion in our method. Our method processes the bun-
setsu sequence of an input sentence in order from the
begining to as follows:
1. We store the bunsetsu sequence of an input sen-
tence in the input order in the queue, and make
both the shift and swap stacks empty.
2. It is repeated that one of five operators (Shift,
Shift-Comma, Reduce, Reduce-Comma and
Swap) is selected to manipulate the target two
bunsetsus. One of the target two bunsetsus (here-
after, the forward bunsetsu) is the top bunsetsu of
the shift stack. The other (hereafter, the backward
bunsetsu) is the front bunsetsu of the queue if the
swap stack is empty and the top bunsetsu of the
swap stack otherwise. In practice, the choice of
operators is limited by the state of two stacks and
the queue, due to constraints on two aspects of the
algorithm’s behavior and Japanese grammar.
3. The repetition of 2. finishes when both the queue
and the swap stack become empty and only a tree
of which the root is the end-of-sentence bunsetsus
is left in the shift stack.
Each operator is elaborated as follows:
Shift operator ensures that the forward bunsetsu does
not depends on the backward one, and moves the
backward one into the shift stack.
Shift-Comma operator performs similar to the Shift
operator and also inserts a comma between the
forward and backward bunsetsus.
Reduce operator specifies that the forward bunsetsu
depends on the backward bunsetsu, removes the
forward bunsetsu from the shift stack, and adds it
as a first child node of the backward bunsetsu to
form a dependency tree.
Reduce-Comma operator performs the same oper-
ation of the Reduce operator. It also inserts a
comma between the forward and backward bun-
setsus.
Swap operator determines to swap the order of the
forward and backward bunsetsus, and pushes
them into the swap stack in this order. By using
the swap stack, we can reset the previous decision
on dependency parsing and comma insertion re-
lated to the target two bunsetsus, and reperforms
these processes based on the swapped word order
again.
We describe a concrete flow of our algorithm pre-
sented in Figure 2. As can be seen in Figure 2, a box
means a bunsetsu and boxes of the target two bunset-
sus are shown with a bold frame. In the initial state
at time 1, the bunsetsu sequence of the input sentence
is stored in the queue, and both the shift stack and the
swap stack are empty, so only the front of the queue
is targeted, and Shift is selected. As a result, “
1
I”
is pushed into the shift stack. At time 2, Shift is se-
lected, and “
2
home” is pushed into the shift stack
because it assumes that there is not a dependency re-
lation and a comma between “
1
I” and “
2
home”.
At time 4, Reduce operator is chosen and “
3
the
city” is removed from the shift stack because there
is not a comma but a dependency relation between
“
3
the city” and “
4
longing for”. At time 5, Swap
is selected, and therefore “
2
home”, “
3
the city” and
“
4
longing for” are pushed into the swap stack in the
order of “
2
home”, “
4
longing for”, “
3
the city”. At
time 6, Shift-Comma is selected, and consequently a
comma is inserted after “
1
I” and “
3
the city,” which
is the top of the swap stack, is pushed into the shift
stack since there is no dependency relation, but then
again a comma between “
1
I” and “
3
the city”. Fi-
nally, at time 14, the queue and the swap stack are
empty, and only the final bunsetsu is on the shift stack,
so the process ends.
3.2 Probabilistic Model
In this section, we describe a probabilistic model used
in the operator selection in our proposed algorithm. In
this study, we conducted a probabilistic model that es-
timates the validity of the processing results generated
by each operator, and selects each operator based on
the highest value of the processing results.
In the following equation, f
t
represents an opera-
tor that has been selected at time t and f
f
f
t
= f
1
f
2
··· f
t
indicates a sequence of operations from time 1 to
time t. B = b
1
b
2
···b
n
defines the bunsetsu sequence
of an input sentence. b
i
distinguishes the ith bunsetsu
in the word order of an input sentence. S
f
f
f
t
presents
the structure expressing the result that f
f
f
t
(operations
up to time t) are performed. The structure S
f
f
f
t
is
defined as a tuple S
f
f
f
t
= hO
f
f
f
t
,C
f
f
f
t
, D
f
f
f
t
i, where O
f
f
f
t
=
{o
f
f
f
t
1,2
, o
f
f
f
t
2,3
, ··· , o
f
f
f
t
1,n
, o
f
f
f
t
2,3
, ··· , o
f
f
f
t
i, j
, ··· , o
f
f
f
t
n−1,n
}, C
f
f
f
t
=
{c
f
f
f
t
1,2
, c
f
f
f
t
2,3
, ··· , c
f
f
f
t
1,n
, c
f
f
f
t
2,3
, ··· , c
f
f
f
t
i, j
, ··· , c
f
f
f
t
n−1,n
}, and
D
f
f
f
t
= {d
f
f
f
t
1,2
, d
f
f
f
t
2,3
, ··· , d
f
f
f
t
1,n
, d
f
f
f
t
2,3
, ··· , d
f
f
f
t
i, j
, ··· , d
f
f
f
t
n−1,n
}
are the word order, the comma positions, and the
dependency structure, respectively, which were
determined by f
f
f
t
.
Here, o
f
f
f
t
i, j
(1 ≤ i < j ≤ n) expresses the order be-
tween b
i
and b
j
after an operation at time t, and o
f
f
f
t
i, j
is
1 if b
i
is located before b
j
after a operation f
f
f
t
, and is 0
otherwise. In addition, c
f
f
f
t
i, j
(1 ≤ i < j ≤ n) is 1 if there
is a comma between b
i
and b
j
, and is 0 otherwise. Fi-
Word Reordering and Comma Insertion Integrated with Shift-Reduce Dependency Parsing
1147
Table 1: Experimental Results.
word reordering comma insertion
pair complete recall precision f-measure
72.61% 10.90% 48.44% 60.78% 53.91
(24,209/33,343)
(109/1,000) (31/64) (31/51)
nally, d
f
f
f
t
i, j
(1 ≤ i < j ≤ n) expresses the dependency
relation between b
i
and b
j
after an operation at time t,
and d
f
f
f
t
i, j
is 1 if b
i
depends on b
j
, and is 0 otherwise.
In the proposed method, the selection of the op-
eration f
t
is calculated by Equation (1). The forward
bunsetsu at time t is denoted by b
i
and the backward
bunsetsu is identified by b
j
, and the difference be-
tween S
f
f
f
t−1
f
t
and S
f
f
f
t−1
is o
f
f
f
t−1
f
t
i, j
, c
f
f
f
t−1
f
t
i, j
, d
f
f
f
t−1
f
t
i, j
.
argmax
f
t
P(S
f
f
f
t−1
f
t
|B, S
f
f
f
t−1
)
≈ argmax
f
t
P(o
f
f
f
t−1
f
t
i, j
, c
f
f
f
t−1
f
t
i, j
, d
f
f
f
t−1
f
t
i, j
|B, S
f
f
f
t−1
)
= argmax
f
t
P(o
f
f
f
t−1
f
t
i, j
|B, S
f
f
f
t−1
)
× P(c
f
f
f
t−1
f
t
i, j
|B, S
f
f
f
t−1
, o
f
f
f
t−1
f
t
i, j
)
× P(d
f
f
f
t−1
f
t
i, j
|B, S
f
f
f
t−1
, o
f
f
f
t−1
f
t
i, j
, c
f
f
f
t−1
f
t
i, j
)
(1)
P(o
f
f
f
t−1
f
t
i, j
|B, S
f
f
f
t−1
), P(c
f
f
f
t−1
i, j
|B, S
f
f
f
t−1
, o
f
f
f
t−1
f
t
i, j
) and
P(d
f
f
f
t−1
|B, S
f
f
f
t−1
, o
f
f
f
t−1
f
t
i, j
, c
f
f
f
t−1
f
t
i, j
) are all estimated by
machine learning. In estimating P(o
f
f
f
t−1
f
t
i, j
|B, S
f
f
f
t−1
),
we used all of the features proposed by (Yoshida et al.,
2014) except for the features related to the bunsetsu
that are depended to the forward bunsetsu or back-
ward one. In estimating P(c
f
f
f
t−1
f
t
i, j
|B, S
f
f
f
t−1
, o
f
f
f
t−1
f
t
i, j
),
we applied all of the features suggested by (Murata
et al., 2010) that can be acquired without the depen-
dency information about the forward bunsetsu. In es-
timating P(d
f
f
f
t−1
f
t
i, j
|B, S
f
f
f
t−1
, o
f
f
f
t−1
f
t
i, j
, c
f
f
f
t−1
f
t
i, j
), all of the
features recommended by (Sassano, 2004) were con-
fidered.
4 EXPERIMENT
To evaluate the performance of the proposed method
for word reordering and comma insertion in hard-to-
read Japanese sentences, we conducted an experiment
using newspaper articles.
4.1 Creation of Test Data
According to the assumption that sentences in news-
paper articles are written in easy-to-read word order,
we created 1000 sentences with hard-to-read word
order based on the following procedure: (1) Create
pseudo-sentences with hard-to-read word order from
a newspaper article (Yoshida et al., 2014). (2) Manu-
ally add commas to make it as easy as possible to read
in that word order.
4.2 Outline of Experiment
We employed a gradient-boosting machine (GBM) as
the machine learning model for estimating each prob-
ability in Equation (1) and used the default Light-
GBM
2
. A total of 35,404 sentences from Kyoto Uni-
versity Text Corpus Ver. 4.0 (Kawahara et al., 2002)
were selected for the training, excluding the sentences
used to create the test data for Section 4.1.
In the evaluation of word reordering, we obtained
the following two measurements, which are defined
by (Uchimoto et al., 2000). Complete agreement is
the percentage of the output sentences in which the
word order entirely agrees with that of the original
sentence. Pair agreement is the percentage of the
pairs of bunsetsus whose word order agrees with the
word order in the original sentence.
For the comma insertion, we evaluated only sen-
tences of which the word order completely agrees.
We measured the precision and recall based on the
assumption that the comma positions in Kyoto Uni-
versity Text Corpus are correct, as with (Murata et al.,
2010).
4.3 Experimental Results
Table 1 presents the pair and complete agreements of
our method for word reordering. It also shows the
recall, precision, and F-measure for the comma inser-
tion in the output sentences whose word order is com-
pletely consistent with the correct answer. Although
the performance of word reordering and comma in-
sertion was not necessarily sufficient, there existed 65
sentences in which the word order and comma posi-
tions matched perfectly between the output sentences
and the correct answers as shown in Figure 3. There-
fore, it can conclude that the feasibility of our method
was confirmed.
2
https://lightgbm.readthedocs.io/en/latest/
ICAART 2021 - 13th International Conference on Agents and Artificial Intelligence
1148
7KHVHQWHQFHLVWUDQVODWHG DVn7KHUHLVDVWURQJWHQGHQF\WRDYRLG
GDQJHUDQGFRPSOHWHIRUYLFWRU\RUGHIHDWZLWKDORQJEDOO|LQ(QJOLVK
ᶅ
ঋഌΝ
YLFWRU\
RUGHIHDW
ᶈ
ϫϱήΫρέͲ
ZLWKDORQJ
EDOO
ᶆ
ثݧΝ
GDQJHU
ᶇ
ඈ͜
WRDYRLG
ᶉ
૬ͮͱ͏Ζ
FRPSHWH
IRU
ɼ
,QSXW
2XWSXW FRUUHFW ZRUG RUGHU FRPPDV GHSHQGHQFLHV
ᶊ
ܑ͗
WHQGHQF\
ᶋ
ک͏ɽ
VWURQJ
ᶅ
ঋഌΝ
YLFWRU\
RUGHIHDW
ᶈ
ϫϱήΫρέͲ
ZLWKDORQJ
EDOO
ᶆ
ثݧΝ
GDQJHU
ᶇ
ඈ͜
WRDYRLG
ᶉ
૬ͮͱ͏Ζ
FRPSHWH
IRU
ᶊ
ܑ͗
WHQGHQF\
ᶋ
ک͏ɽ
VWURQJ
Figure 3: Successful example of word ordering, comma insertion, and dependency parsing.
5 CONCLUSION
This paper proposed a method of simultaneously de-
termining appropriate word order, appropriate comma
positions, and the dependency structure for an in-
put sentence which has unsuitable word sequence.
The proposed algorithm in this study is further ex-
tended the Shift-Reduce algorithm suggested (Sas-
sano, 2004) and focused on two local bunsetsus. We
confirmed the feasibility of our method as the results
of the evaluation experiment using 1000 sentences
with hard-to-read word order.
For future works, we are interested in improving
the agreements on word reordering, and the precision
and recall on comma insertion by reviewing the fea-
tures and the machine learning method used for prob-
ability estimation.
ACKNOWLEDGMENTS
This work was partially supported by JSPS KAK-
ENHI Grant Numbers JP26280082, JP16K00300,
JP19K12127.
REFERENCES
Belz, A. and Kow, E. (2011). Discrete vs. continuous rating
scales for language evaluation in NLP. In Proceed-
ings of the 49th Annual Meeting of the Association for
Computational Linguistics: Human Language Tech-
nologies, pages 230–235.
Filippova, K. and Strube, M. (2007). Generating constituent
order in German clauses. In Proceedings of the 45th
Annual Meeting of the Association of Computational
Linguistics, pages 320–327.
Harbusch, K., Kempen, G., van Breugel, C., and Koch, U.
(2006). A generation-oriented workbench for perfor-
mance grammar: Capturing linear order variability in
German and Dutch. In Proceedings of the 4th Inter-
national Natural Language Generation Conference,
pages 9–11.
Kawahara, D., Kurohashi, S., and Hasida, K. (2002). Con-
struction of a Japanese relevance-tagged corpus. In
Proceedings of the 3rd International Conference on
Language Resources and Evaluation, pages 2008–
2013.
Kruijff, G.-J. M., Kruijff-Korbayov´a, I., Bateman, J., and
Teich, E. (2001). Linear order as higher-level deci-
sion: Information structure in strategic and tactical
generation. In Proceedings of the 8th European Work-
shop on Natural Language Generation, pages 74–83.
Murata, M., Ohno, T., and Matsubara, S. (2010). Automatic
comma insertion for Japanese text generation. In Pro-
ceedings of the 2010 Conference on Empirical Meth-
ods in Natural Language Processing, pages 892–901.
Ringger, E., Gamon, M., Moore, R. C., Rojas, D., Smets,
M., and Corston-Oliver, S. (2004). Linguistically in-
formed statistical models of constituent structure for
ordering in sentence realization. In Proceedings of the
20th International Conference on Computational Lin-
guistics, pages 673–679.
Sassano, M. (2004). Linear-time dependency analysis for
Japanese. In Proceedings of the 20th International
Conference on Computational Linguistics, pages 8–
14.
Shaw, J. and Hatzivassiloglou, V. (1999). Ordering among
premodifiers. In Proceedings of the 37th Annual Meet-
ing of the Association for Computational Linguistics,
pages 135–143.
Uchimoto, K., Murata, M., Ma, Q., Sekine, S., and Isahara,
H. (2000). Word order acquisition from corpora. In
Proceedings of the 18th International Conference on
Computational Linguistics, pages 871–877.
Yokobayashi, H., Suganuma, A., and Taniguchi, R.-i.
(2004). Generating candidates for rewriting based on
Word Reordering and Comma Insertion Integrated with Shift-Reduce Dependency Parsing
1149
an indicator of complex dependency and it’s applica-
tion to a writing tool. Journal of Information Process-
ing Society of Japan, 45(5):1451–1459. (In Japanese).
Yoshida, K., Ohno, T., Kato, Y., and Matsubara, S. (2014).
Japanese word reordering integrated with dependency
parsing. In Proceedings of the 25th International Con-
ference on Computational Linguistics: Technical Pa-
pers, pages 1186–1196.
ICAART 2021 - 13th International Conference on Agents and Artificial Intelligence
1150
