Proceedings of the Fifth SIGHAN Workshop on Chinese Language Processing, pages 94–101,
Sydney, July 2006. c©2006 Association for Computational Linguistics
A SVM-based Model for Chinese Functional Chunk Parsing 
 
 
Yingze Zhao 
State Key Laboratory of Intelligent Technol-
ogy and Systems 
Dept. of Computer Science and Technology, 
Tsinghua University 
Beijing 100084, P. R. China  
zhaoyingze@gmail.com 
Qiang Zhou 
State Key Laboratory of Intelligent Technol-
ogy and Systems 
Dept. of Computer Science and Technology, 
Tsinghua University 
Beijing 100084, P. R. China 
zq-lxd@mail.tsinghua.edu.cn 
 
  
 
Abstract 
Functional chunks are defined as a series 
of non-overlapping, non-nested segments 
of text in a sentence, representing the im-
plicit grammatical relations between the 
sentence-level predicates and their argu-
ments. Its top-down scheme and com-
plexity of internal constitutions bring in a 
new challenge for automatic parser. In 
this paper, a new parsing model is pro-
posed to formulate the complete chunk-
ing problem as a series of boundary de-
tection sub tasks. Each of these sub tasks 
is only in charge of detecting one type of 
the chunk boundaries. As each sub task 
could be modeled as a binary classifica-
tion problem, a lot of machine learning 
techniques could be applied.  
In our experiments, we only focus on 
the subject-predicate (SP) and predicate-
object (PO) boundary detection sub tasks. 
By applying SVM algorithm to these sub 
tasks, we have achieved the best F-Score 
of 76.56% and 82.26% respectively. 
1 Introduction 
Parsing is a basic task in natural language proc-
essing; however, it has not been successful in 
achieving the accuracy and efficiency required 
by real world applications. As an alternative, 
shallow parsing or partial parsing has been pro-
posed to meet the current needs by obtaining 
only a limited amount of syntactic information 
needed by the application. In recent years, there 
has been an increasing interest in chunk parsing. 
From CoNLL-2000 to CoNLL-2005, a lot of ef-
forts have been made in the identification of ba-
sic chunks and the methods of combining them 
from bottom-up to form large, complex units. In 
this paper, we will apply functional chunks to 
Chinese shallow parsing. 
Functional chunks are defined as a series of 
non-overlapping, non-nested functional units in a 
sentence, such as subjects, predicates, objects, 
adverbs, complements and so on. These units 
represent the implicit grammatical relations be-
tween the sentence-level predicates and their ar-
guments. Different from the basic chunks de-
fined by Abney (1991), functional chunks are 
generated from a top-down scheme, and thus 
their constitutions may be very complex. In addi-
tion, the type of a functional chunk could not be 
simply determined by its constitution, but de-
pends heavily on the context. Therefore, we will 
have new challenges in the functional chunk 
parsing. 
Ramshaw and Marcus (1995) first introduced 
the machine learning techniques to chunking 
problem. By formulating the NP-chunking task 
as a tagging process, they marked each word 
with a tag from set {B, I, O}, and successfully 
applied TBL to it. Inspired by their work, we 
introduce SVM algorithm to our functional 
chunking problem. Instead of using the BIO tag-
ging system, we propose a new model for solv-
ing this problem. In this model, we do not tag the 
words with BIO tags, but directly discover the 
chunk boundaries between every two adjacent 
functional chunks. Each of these chunk bounda-
ries will be assigned a type to it, which contains 
the information of the functional chunk types 
before and after it. Then we further decompose 
this model into a series of sub modules, each of 
which is in charge of detecting only one type of 
94
the chunk boundaries. As each sub module can 
be modeled as a binary classifier, various ma-
chine learning techniques could be applied. 
In our experiments, we focus on the subject-
predicate (SP) and predicate-object (PO) bound-
ary detection tasks, which are the most difficult 
but important parts in our parsing model. By ap-
plying SVM algorithm to these tasks, we achieve 
the best F-Score of 76.56% and 82.26% respec-
tively.                                                                                                              
This paper is organized as follows. In section 
2, we give a brief introduction to the concept of 
our functional chunks. In section 3, we propose 
the parsing model for Chinese functional chunk 
parsing. In section 4, we compare SVM with sev-
eral other machine learning techniques, and illus-
trate how competitive SVM is in our chunking 
task. In section 5, we build 2 sub modules based 
on SVM algorithm for SP and PO boundary de-
tection tasks. In section 6, some related work on 
functional chunk parsing is introduced. Section 7 
is the conclusion. 
2 Functional Chunk Scheme 
Functional chunks are defined as a series of 
non-overlapping, non-nested segments of text at 
the sentence level without leaving any words 
outside. Each chunk is labeled with a functional 
tag, such as subject, predicate, object and so on. 
These functional chunks in the sentence form a 
linear structure within which the grammatical 
relations between sentence-level predicates and 
their arguments or adjuncts are kept implicitly. 
Table 1 lists all the tags used in our functional 
chunk scheme:  
Table 1. Functional Chunk Tag Set. 
Chunk Tag Basic Function Description
S Subject 
P Predicate 
O Object 
J Raised Object 
D Adverbial adjunct 
C Complement 
T Independent constituent 
Y Modal particle 
Here, we list some examples to illustrate how 
these functional tags are used in Chinese sen-
tences. 
1. “[D /� /t (afternoon)  /  [D � /p 
(when)  � /rN (I) 	�� /v (come to)  ��h
/nS (Xi Bai Po village) �	g /s (eastern entrance)  
H /n  /  [D X /d (already) [P � /v (there is) 
[J B /m  � /qN (a) � /n (brainman) [D  /p  
*	� /rS (there) [P �� /v (waiting) [Y 
 /y  b
/ b  “ 
2. “[T 	V[� /l (frankly speaking)  /  [S 
* /rN (that) [P ^ /vC (was) [O � /rN (I) �3 /d 
(lifetime) � /dN  ? /vM (can’t) �� /v (forget) 
� /u   b/ b  “ 
3. “[S HW /n (time) [P �� /v  � /u (schedule) 
[C � /dD (very) � /a (tight)  b/ b  “ 
Compared with the basic chunk scheme de-
fined by Abney (1991), our functional chunk 
scheme has the following two main differences: 
(1) Functional chunks are not constituted from 
bottom-up, but generated from top-down, thus 
some functional chunks are usually longer and 
more complex than the basic chunks.  
We have a collection of 185 news files as our 
functional chunk corpus. Each file is manually 
annotated with functional chunks. There are 
about 200,000 Chinese words in the corpus. To 
investigate the complex constitutions of func-
tional chunks, we list the average chunk lengths 
(ACL) of different types in Table 2: 
Table 2. Average Chunk Lengths of Different 
Types. 
Chunk Type Count Word Sum ACL
P 21988 27618 1.26 
D 19795 46919 2.37 
O 14289 61401 4.30 
S 11920 34479 2.89 
J 855 2083 2.44 
Y 594 604 1.02 
T 407 909 2.23 
C 244 444 1.82
From the table above, we can find that O 
chunk has the longest average length of 4.30 
words, and S chunk has the second longest aver-
age length of 2.89 words, and D chunk has an 
average length of 2.37 words. Although the aver-
age length doesn’t seem so long, the length of a 
specific chunk varies greatly.  
In Table 3, we list some detailed length distri-
butional data of three chunks. 
Table 3. Length Distribution of S, O and D 
Chunks. 
Chunk Length # of S # of O # of D
1 5322 3537 12147
2 2093 2228 2499 
3 1402 2117 1431 
4 917 1624 1010 
5 627 1108 696 
>5 1559 3675 2013 
Sum 11920 14289 19796
95
From the table above, we can find that there 
are totally 1559 S chunks with a length of more 
than 5 words which takes up 13.08% of the total 
number. And when we refer to the S chunks with 
more than 3 words, the percentage will increase 
to 26.03%. These long chunks are usually consti-
tuted with several complex phrases or clauses as 
the modifiers of a head word. Among the O 
chunks, 25.72% of them have a length of more 
than 5 words, and 44.84% of them are longer 
than 3 words. The reason why O chunks have a 
longer length may be that many of them contain 
the entire clauses. Although most of the D 
chunks are less than 5 words, some constituted 
with complex preposition phrases can still be 
very long. 
The complex constitutions of S, O, D chunks 
are the main parsing difficulties. 
(2) The type of functional chunks can’t be 
simply determined by their constitutions, but de-
pends heavily on their contexts. 
As the constitution of a basic chunk is very 
simple, its type can be largely determined by its 
head word, but in the case of functional chunks, 
the relationships between the functional chunks 
play an important role. For example, a NP phrase 
before a P chunk can be identified as a subject 
chunk, but in other sentences, when it follows 
another P chunk, it will be recognized as an ob-
ject chunk. Thus we can’t determine the type of a 
functional chunk simply by its constitution. 
The context dependencies of functional 
chunks bring a new challenge for our chunk 
parser. 
In the next section, we will propose a top-
down model for Chinese functional chunk pars-
ing. Since the functional chunk boundaries have 
the information of linking two adjacent chunks, 
they will be very helpful in the determination of 
chunk types. 
3 Parsing Model 
The Chinese functional chunk parser takes a 
stream of segmented and tagged words as its in-
put, and outputs all the functional chunk bounda-
ries in a sentence. In this section, we will present 
a parsing model which formulates the functional 
chunk parsing problem as a boundary detection 
task, and then decompose this model into a series 
of sub modules that are easy to build. 
3.1 Formulation 
Functional chunks have the property of exhaust-
ibility and no words will be left outside the 
chunks. Thus we don’t need to find the end posi-
tion for a functional chunk as it could be identi-
fied by the start of the next one. In this case, we 
can simply regard the chunking task as a process 
of cutting the input sentence into several seg-
ments of words, each of which is labeled with a 
functional tag. Based on this idea, we can model 
the functional chunk parsing problem as a 
boundary detection task. 
Let S=<W, T> denote the input sentence to be 
parsed by the functional chunk parser, where 
W=w
1
w
2
w
3
…w
n
 is the sequence of words in S, 
and T=t
1
t
2
t
3
…t
n
 is sequence of the POS tags as-
signed to each word in W. If w
i
 is a punctuation 
mark, t
i
 will be equal to w
i
. 
A chunk boundary is defined as a pair <C
1
, 
C
2
> where  C
1
 ,C
2 
 �{S, P, O, J, D, C, T, Y}, C
1
 
is the chunk type before this boundary and C
2
 is 
the chunk type following it. The output of the 
chunk parser is denoted as O=<B, P> where 
B=b
1
b
2
b
3
…b
m
 is the sequence of chunk bounda-
ries generated by the parser, and P=p
1
p
2
p
3
…p
m
 is 
the corresponding positions of b
1
b
2
b
3
…b
m
 in the 
sentence. 
Chinese functional chunk parser can be con-
sidered as a function h(S) which maps the input 
sentence S to the chunk boundary sequence O.  
Take the following sentence for example: 
“14  �� /n(Nuclear electricity) 
1
 ^ /vC(is) 
2
 
B /m(a) 
 3
 � /qN(kind) 
4
 �� /a(safe) 
5
  a/ a 
6
  
b� /a(safe) 
7
  a/ a 
8
 �6 /a(economical) 
9
 � /u 
10
  ?� /n(energy) 
11
  b/ b” 
“Nuclear electricity is a kind of safe, clean and 
economical energy.” 
In this sentence, there are totally 12 Chinese 
words (punctuation marks are treated the same 
way as words) with 11 numbers falling between 
them indicating the positions where a functional 
chunk boundary may appear. If the input sen-
tence is parsed correctly by the functional chunk 
parser, a series of boundaries will arise at posi-
tion 1 and 2, which are illustrated as below: 
“14  �� /n 
<S, P>
 ^ /vC 
<P, O>
 B /m � /qN �
� /a  a/ a  b� /a  a/ a  �6 /a � /u  ?� /n  b
/ b” 
 From the information provided by these 
boundaries, we can easily identify the functional 
chunks in the sentence: 
“14  [S �� /n  [P ^ /vC  [O B /m  � /qN  �
� /a   a/ a  b� /a   a/ a  �6 /a  � /u  ?�
/n   b/ b” 
96
3.2 Decomposition of Parsing Model 
The functional chunk parser presented above 
could be further divided into several sub modules, 
each of which is only in charge of detecting one 
type of the chunk boundaries in a sentence. The 
sub module in charge of detecting boundary b 
could be formulated as a Boolean function h
b
(S, i) 
where S is the input sentence and i is the position 
between word w
i
 and w
i+1
. Function h
b
(S, i) will 
take true if there is a chunk boundary of type b at 
position i, and it will take false if there’s not. 
Since the Boolean function h
b
(S, i) can be treated 
as a binary classifier, many machine learning 
techniques could be applied. 
If we combine every two chunk types in the 
tag set, we can make a total number of 8*8=64 
boundary types in our chunking task. However, 
not all of them appear in the natural language 
text, for example, we don’t have any SO bounda-
ries in our corpus as S and O chunks can’t be-
come neighbors in a sentence without any P 
chunks between them. In our corpus, we could 
find 43 boundary types, but only a small number 
of them are used very frequently. In table 4, we 
list the 5 most frequently used boundaries in our 
corpus:  
Table 4. The 5 Most Frequently Used Bounda-
ries in the Corpus. 
Boundary Type Count 
PO 14209 
DP 11459 
SD 6156 
DD 5238 
SP 5233 
The top 5 boundaries take up 67.76% of all the 
62418 boundaries in our corpus. If we further 
investigate the chunk types associated with these 
boundaries, we can find that only four types are 
involved: P, D, O and S. Referred to Table 2, we 
can find that these chunks are also the 4 most 
frequently used chunks in our corpus. 
In most cases, S, P, and O chunks constitute 
the backbone of a Chinese sentence, and they 
usually contain the most useful information we 
need. Therefore, we are more concerned about S, 
P and O chunks. In the following sections, we 
will focus on the construction of sub modules for 
SP and PO boundary detection tasks. 
4 Statistical Model Selection 
After decomposing the parsing model into sev-
eral sub modules, a lot of machine learning tech-
niques could be applied to the constructions of 
these sub modules. 
SVM
1
 is a machine learning technique for 
solving the binary classification problems. It is 
well known for its good generalization perform-
ance and high efficiency. In this section, we will 
make a performance comparison between SVM  
(Vapnik, 1995) and several other machine learn-
ing techniques including Naïve Bayes, ID3
2
 
(Quinlan, 1986) and C4.5
3
 (Quinlan, 1993), and 
then illustrates how competitive SVM is in the 
boundary detection tasks. 
4.1 Experimental Data 
The corpus we use here is a collection of 185 
news files which are manually corrected after 
automatic sentence-split, word segmentation and 
part-of-speech tagging. After these processes, 
they have been manually annotated with func-
tional chunks. Among the 185 files, 167 of them 
are taken as the training data and the remaining 
18 are left as the test data, which takes up ap-
proximately 10% of all the data.  
In our experiments, we will use feature tem-
plates to describe which features are to be used 
in the generation of feature vectors. For example, 
if the current feature template we use is w-1t2, 
then the feature vector generated at position i 
will take the first word on the left and the second 
word tag on the right as its features. 
Before we perform any experiments, all the 
data have been converted to the vectors that are 
acceptable by different machine learning algo-
rithms. Thus we have a total number of 199268 
feature vectors generated from the 185 files. 
Among them, 172465 vectors are in the training 
data and 26803 vectors are in the test data. Two 
sets of training and test data are prepared respec-
tively for the SP and PO boundary detection 
tasks.  
The performance of each experiment is meas-
ured with 3 rates: precision, recall and F
β =1
, 
where precision is the percentage of detected 
boundaries that are correct, recall is the percent-
age of boundaries in the test data that are found 
by the parser, and F
β =1 
is defined as 
F
β
=(β
2
+1)*precision*recall/(β
2
*precision + recall) 
with β =1.  
                                                 
1
 The software package we use is SVM
light
 v6.00, it is avail-
able at http://svmlight.joachims.org/. We use linear kernel 
function and other default parameters in our experiments. 
2
 We use the weka’s implementation of Naïve Bayes and 
ID3 algorithms. Weak 3.4 is available at 
http://www.cs.waikato.ac.nz/ml/weka/. 
3
 We use Quinlan’s C4.5 software package with its default 
parameters in our experiments. 
97
4.2 Algorithm Comparison 
We first use t-3t-2t-1t1t2 as the feature tem-
plate, and list all the experimental results in Ta-
ble 5 and Table 6. From these results, we can 
find that SVM has achieved the best precision, 
recall and F-Score in SP boundary detection task, 
while C4.5 has an overwhelming advantage in 
PO boundary detection task. In both tasks, Naïve 
Bayes algorithm performs the worst, which 
makes us very disappointed. 
Table 5. Results of Different Algorithms in SP 
Boundary Detection Task. 
Algorithm Precision Recall F
β =1
 
SVM 82.21% 57.10% 67.39%
ID3 67.60% 50.70% 57.94%
C4.5 81.10% 44.60% 57.55%
Naïve Bayes 47.90% 51.00% 49.40%
Table 6. Results of Different Algorithms in 
PO Boundary Detection Task. 
Algorithm Precision Recall F
β =1
 
C4.5 72.00% 74.70% 73.33%
SVM 67.27% 64.96% 66.09%
ID3 70.70% 59.90% 64.85%
Naïve Bayes 48.10% 60.10% 53.43%
As the feature template we use here is too sim-
ple, the results we have got may not seem so per-
suasive. Therefore we decide to conduct another 
experiment using a more complex feature tem-
plate.  
In the following experiments, we will use w-
2w-1w1w2t-2t-1t1t2 as the feature template. The 
experimental results are listed in Table 7 and Ta-
ble 8. 
After adding the word information to the fea-
ture template, the dimensions of feature vectors 
used by some algorithms increase dramatically. 
We remove Naïve Bayes algorithm from the fol-
lowing experiments, as it fails to deal with such 
high dimensional data.  
Table 7. Results of Different Algorithms in SP 
Boundary Detection Task. 
Algorithm Precision Recall F
β =1
 
SVM 82.25% 61.22% 70.19%
ID3 64.70% 51.70% 57.47%
C4.5 79.70% 37.40% 50.91%
Table 8. Results of Different Algorithms in 
PO Boundary Detection Task. 
Algorithm Precision Recall F
β =1
 
SVM 74.83% 86.99% 80.45%
C4.5 67.90% 79.90% 73.41%
ID3 75.10% 57.70% 65.26%
After applying the complex feature template, 
SVM still keeps the first place in SP boundary 
detection task. In PO boundary detection task, 
SVM successfully takes the place of C4.5, and 
achieves the best recall and F-Score among all 
the algorithms. Although the precision of ID3 is 
a little better than SVM, we still prefer SVM to 
ID3. It seems that the word information in the 
feature vectors is not so beneficial to decision 
tree algorithms as to SVM.  
We also notice that SVM can perform very ef-
ficiently even with a large number of features. In 
the second set experiments, it usually takes sev-
eral hours to train a decision tree model, but for 
SVM, the time cost is no more than 20 minutes. 
In addition, we can expect a better result by add-
ing more information to SVM algorithm without 
worrying about the dimension disaster problem 
in other algorithms. Therefore, we decide to base 
our parsing model on SVM algorithm. 
5 The SVM-based Parsing Model 
5.1 Baseline Models 
In this section, we will build 2 baseline models 
based on SVM for SP and PO boundary detec-
tion tasks respectively. By comprising the results 
of two different feature templates, we will illus-
trate how useful the word information is in our 
SVM based models.  
One feature template we use here is the simple 
template which only takes the POS tag informa-
tion as its features. The other one is the complex 
template which takes both word and tag informa-
tion as its features. To make sure the results are 
comparable, we restrict the context window to 4 
words. 
In the SP boundary detection sub task, we got 
the following results: 
Table 9. SP Boundary Detection Results.  
Feature template Precision Recall F
β =1
 
t-2t-1t1t2 76.25% 51.99% 61.83%
w-2w-1w1w2t-
2t-1t1t2 
82.25% 61.22% 70.19%
In the PO boundary detection sub task, we got 
the following results: 
Table 10. PO Boundary Detection Results.  
Feature template Precision Recall F
β =1
 
t-2t-1t1t2 66.42% 65.27% 65.84%
w-2w-1w1w2t-
2t-1t1t2 
74.83% 86.99% 80.45%
 
By taking the complex feature template, we 
have achieved the best F
β =1
 value of 70.19% in 
SP boundary detection experiment and 80.45% 
in PO experiment, both of which are much 
higher than those of the simple feature templates. 
From these results we can conclude that word 
information is very helpful in our SVM based 
98
models. Thus we will only use the feature tem-
plates with word information in the succeeding 
experiments. 
5.2 Expanding the Context Window 
In the previous section, the feature templates we 
use are restricted to a context window of 4 words, 
which might not be large enough to detect the 
boundaries between complex chunks. For exam-
ple, when parsing the sentence “[P � /v 
1
 [O 
�� /a 
2
 � /u 
3
 
�B /n 
4
 �� /n 
5
 �. /vN 
6
 ��
/n”, the algorithm fails to detect the PO boundary 
at position 1. If we expand the context window to 
the noun word “
�B /n”, some of these errors 
may disappear. In the following experiments, we 
will expand the context window from a size of 4 
words to 10 words, and make a comparison be-
tween the different results. 
The 4 feature templates used here are listed 
below: 
T1: w-2w-1w1w2t-2t-1t1t2,  
T2: w-3w-2w-1w1w2w3t-3t-2t-1t1t2t3,  
T3: w-4w-3w-2w-1w1w2w3w4t-4t-3t-2t-
1t1t2t3t4 
T4: w-5w-4w-3w-2w-1w1w2w3w4w5t-5t-4t-
3t-2t-1t1t2t3t4t5.  
 4 1  # P V O E B S Z  % F U F D U J P O  3 F T V M U T
      
      
             
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
 5   5   5   5 
 1 S F D J T J P O  3 F D B M M  '  4 D P S F
  
   Figure 1. SP Boundary Detection Results. 
As we have expected, the performance of SP 
boundary detection experiment has been im-
proved as the context window expands from a 
size of 4 words to 8 words. However, the preci-
sion value meets its turning point at T3 after 
which it goes down, while F-Score and recall 
value still keep rising. From the curves shown in 
figure 1, we can find that the expansion of con-
text window size from 4 words to 6 words has an 
obvious improvement for performance, and after 
that only F-Score and recall could be improved.  
 1 0  # P V O E B S Z  % F U F D U J P O  3 F T V M U T
      
      
      
      
      
      
             
      
      
      
      
      
      
      
      
      
      
      
      
      
 5   5   5   5 
 1 S F D J T J P O  3 F D B M M  '  4 D P S F
 
Figure 2. SP Boundary Detection Results. 
In contrast to the significant improvement we 
have achieved in the SP experiments, the results 
of PO experiments are not so exciting. As the 
context window expands, the precision value 
keeps rising while the recall value keeps declin-
ing. Fortunately, we have obtained a very slight 
increase of F-Score from these efforts.  
Although it is very difficult to improve the 
performance of PO boundary detection by simply 
expanding the context window, we’ve still got a 
better result than that of SP. If we examine the 
results of the two tasks carefully, we can find a 
very interesting difference between them: in SP 
boundary detection task, it’s very easier to get a 
better precision than recall, but in PO experiment, 
as the O chunks have a longer length, they are 
more likely to be cut into small pieces, and thus 
it’s easier to get a better recall than precision. 
5.3 Error Analysis 
In our experiments, the recall value can be sim-
ply raised by adding a positive bias value to the 
SVM classifier. However, we can’t do the same 
thing to improve the precision value. Thus, in the 
following analysis, we are only focus on the er-
rors that deter the improvement of precision 
value. 
There are 2 kinds of errors influencing the 
precision value of the test results: One is the 
wrongly detected chunk boundaries (WDB) 
within chunks (these chunk boundaries are de-
tected by the program, but they don’t exist in the 
training data). This kind of error tends to cut a 
large chunk into several small pieces. The other 
is the misclassification of chunk boundary types 
(MBT) at the chunk boundaries (There exists a 
99
chunk boundary at that position, but chunk 
boundary type labeled by the program is wrong). 
 In the following analysis, by comparing the 
numbers of errors in the test results of T1 (w-2w-
1w1w2t-2t-1t1t2) and T4 (w-5w-4w-3w-2w-
1w1w2w3w4w5t-5t-4t-3t-2t-1t1t2t3t4t5), we 
will point out which kind of errors could be ef-
fectively eliminated by the expansion of context 
window and which of them couldn’t. Through 
this analysis, we hope to get some knowledge of 
what efforts should be made in our further study.  
In SP boundary detection task, we list the 
number of wrongly detected chunk boundaries 
(#WDB) and the corresponding chunk types (CT) 
where WDB arises in the following table. 
Table 11. Wrongly Detected Chunk Bounda-
ries in the Test Results of T1 and T4. 
CT #WDB of T1 #WDB of T4 T4-T1
O 17 18 1 
S 17 18 1 
D 7 6 -1 
C 0 1 1 
P 2 1 -1 
T 1 1 0 
Sum 44 45 1 
From the above table, we find that the number 
of wrongly detected boundaries seems to be un-
changed during the expansion of context window. 
But when we refer to the second type of errors, 
the expansion of context window does help. We 
list the misclassified boundary types (MBT) and 
the error numbers (#MB) in the below table. In 
SP boundary detection task, MBT is wrongly 
recognized as boundary type SP. 
Table 12. Misclassified Chunk Boundaries in 
the Test Results of T1 and T4. 
MBT #MB of T1 #MB of T4 T4-T1
OP 9 3 -6 
JP 8 2 -6 
DP 23 20 -3 
SD 6 6 0 
DS 1 1 0 
Sum 47 32 -15 
From the above table, we can find that the 
misclassifications of OP, JP and DP as SP have 
been largely reduced by expanding the context 
window, but the misclassifications of DS and SD 
remain the same. Therefore, we should try some 
other methods for D chunks in our future work. 
In PO boundary detection task, the expansion 
of context window seems to be very effective. 
We list all the results in the below table: 
Table 14. Wrongly Detected Chunk Bounda-
ries in the Test Results of T1 and T4. 
CT #WDB of T1 #WDB of T4 T4-T1
O 251 196 -55 
S 106 76 -30 
D 92 55 -37 
P 56 64 8 
T 4 4 0 
C 1 1 0 
J 0 1 1 
Sum 510 397 -113 
It’s very exciting to see that by expanding the 
window size, the number of WDB decreases dra-
matically from 510 to 397. But it fails to elimi-
nate the WDB errors within P, T, C, and J 
chunks. 
In PO boundary detection task, MBT is 
wrongly recognized as boundary type PO. We 
list the error data of T1 and T4 in the below table. 
Table 13. Misclassified Chunk Boundaries in 
the Test Results of T1 and T4. 
MBT #MB of T1 #MB of T4 T4-T1
PJ 17 18 1 
PD 9 9 0 
PC 8 8 0 
SP 6 6 0 
PS 5 5 0 
SD 5 4 -1 
DP 3 2 -1 
TS 3 3 0 
OD 1 0 -1 
PY 1 1 0 
Sum 58 56 -2 
In contrast to the results of SP boundary detec-
tion task, the MBT errors could not be largely 
reduced by simply expanding the context win-
dow. Therefore, we need to pay more attention to 
these problems in our future work. 
6 Related works 
After the work of Ramshaw and Marcus (1995) , 
many machine learning techniques have been 
applied to the basic chunking task, such as Sup-
port Vector Machines (Kudo and Matsumoto, 
2001), Hidden Markov Model(Molina and Pla 
2002), Memory Based Learning (Sang, 2002), 
Conditional Random Fields (Sha and Pereira, 
2003), and so on. But only a small amount of 
attention has been paid to the functional chunk 
parsing problem. 
Sandra and Erhard (2001) tried to construct 
the function-argument structures based on the 
pre-chunked input. They proposed a similarity 
based algorithm to assign the functional labels to 
complete syntactic structures, and achieved a 
100
precision of 89.73% and 90.40% for German and 
English respectively. Different from our top-
down scheme, their function-argument structures 
are still constituted from bottom-up, and the pre-
chunked input helps simplify the chunking proc-
ess. 
Elliott and Qiang Zhou (2001) used the BIO 
tagging system to identify the functional chunks 
in a sentence. In their experiments, they used 
C4.5 algorithm to build the parsing model, and 
focused their efforts on the selection of feature 
sets. After testing 5 sets of features, they have 
achieved the best f-measure of 0.741 by using 
feature set E which contains all the features in 
other feature sets. Instead of using BIO tags in 
our chunking task, we introduced chunk bounda-
ries to help us identify the functional chunks, 
which could provide more relational information 
between the functional chunks.  
7 Conclusions and Future Works 
In this paper, we have applied functional chunks 
to Chinese shallow parsing. Since the functional 
chunks have the properties of linearity and ex-
haustibility, we can formulate the functional 
chunk parsing problem as a boundary detection 
task. By applying the divide-and-conquer strat-
egy, we have further decomposed the parsing 
model into a series of sub modules, each of 
which is only in charge of one boundary type. In 
this way, we provide a very flexible framework 
within which different machine learning tech-
niques could be applied. In our experiments, we 
build two sub modules based on SVM for solv-
ing the SP and PO boundary detection tasks. 
Thanks to the good generalization performance 
and high efficiency of SVM, we can successfully 
deal with a large number of features. By expand-
ing the context window, we have achieved the 
best F-Score of 76.56% and 82.26 for SP and PO 
boundary detection tasks. 
The 2 sub modules we have built are only 
parts of the Chinese functional chunk parser. Al-
though the results we have got here seem some-
what coarse, they could already be used in some 
simple tasks. In the future, we will build the 
other sub modules for the remaining types of the 
chunk boundaries. After all these work, there 
may be some inconsistent chunk boundaries in 
the results, thus we need to solve the inconsis-
tency problems and try to identify all the func-
tional chunks in a sentence by combining these 
chunk boundaries. 
Acknowledgements 
This work was supported by the Chinese National 
Science Foundation (Grant No. 60573185, 
60520130299).  

References 
Elliott Franco Drábek and Qiang Zhou. 2001. Ex-
periments in Learning Models for Functional 
Chunking of Chinese Text. IEEE International 
Workshop on Natural Language processing and 
Knowledge Engineering, Tucson, Arizona, pages 
859-864. 
E.F. Tjong Kim Sang. 2002. Memory-based shallow 
parsing, Journal of Machine Learning Research 2, 
pages 559-594. 
F. Sha and F. Pereira. 2003. Shallow parsing with 
conditional random fields. In Proceedings of Hu-
man Language Technology Conference / North 
American Chapter of the Association for Computa-
tional Linguistics annual meeting. 
Ian H. Witten and Eibe Frank. 2005. Data Mining: 
Practical machine learning tools and techniques, 
2nd Edition, Morgan Kaufmann, San Francisco.  
Taku Kudo and Yuji Matsumoto. 2001. Chunking 
with support vector machines. Proceedings of the 
Second Meeting of the North American Chapter of 
the Association for Computational Linguistics. 
Pittsburgh, PA. 
Lance Ramshaw and Mitch Marcus. 1995. Text 
chunking using transformation-based learning. In 
Proceedings of the Third Workshop on Very Large 
Corpora, pages 82—94. 
Quinlan, J. Ross. 1986. Induction of decision trees. 
Machine Learning, 1(1), pages 81-106. 
Quinlan, J. Ross. 1993. C4.5: Programs for Machine 
Learning. San Mateo, CA: Morgan Kaufmann. 
Steven Abney. 1991. Parsing by chunks. In Principle-
Based Parsing, Kluwer Academic Publishers, 
Dordrecht, pages 257—278. 
Sandra Kübler and Erhard W. Hinrichs. 2001. From 
chunks to function-argument structure: A similar-
ity-based approach. In Proceedings of ACL/EACL 
2001, Toulouse, France, 2001, pages 338 - 345. 
Thorsten Joachims. 1999. Advances in Kernel Meth-
ods - Support Vector Learning, chapter Making 
large-Scale SVM Learning Practical. MIT-Press. 
Vladimir N. Vapnik. 1995. The Nature of Statistical 
Learning Theory. Springer, New York. 
