Proceedings of the COLING/ACL 2006 Interactive Presentation Sessions, pages 41–44,
Sydney, July 2006. c©2006 Association for Computational Linguistics
Computational Analysis of Move Structures in Academic Abstracts 
Jien-Chen Wu
1
   Yu-Chia Chang
1
   Hsien-Chin Liou
2
   Jason S. Chang
1
 
CS
1
 and FLL
2
, National Tsing Hua Univ. 
{d928322,d948353}@oz.nthu.edu.tw, hcliu@mx.nthu.edu.tw, 
jason.jschang@gmail.com 
Abstract 
This paper introduces a method for 
computational analysis of move 
structures in abstracts of research articles. 
In our approach, sentences in a given 
abstract are analyzed and labeled with a 
specific move in light of various 
rhetorical functions. The method involves 
automatically gathering a large number 
of abstracts from the Web and building a 
language model of abstract moves. We 
also present a prototype concordancer, 
CARE, which exploits the move-tagged 
abstracts for digital learning. This system 
provides a promising approach to Web-
based computer-assisted academic 
writing. 
1 Introduction 
In recent years, with the rapid development of 
globalization, English for Academic Purposes 
has drawn researchers' attention and become the 
mainstream of English for Specific Purposes, 
particularly in the field of English of Academic 
Writing (EAW). EAW deals mainly with genres, 
including research articles (RAs), reviews, 
experimental reports, and other types of 
academic writing. RAs play the most important 
role of offering researchers the access to actively 
participating in the academic and discourse 
community and sharing academic research 
information with one another. 
Abstracts are constantly regarded as the first 
part of RAs and few scholarly RAs go without an 
abstract. “A well-prepared abstract enables 
readers to identify the basic content of a 
document quickly and accurately.” (American 
National Standards Institute, 1979) Therefore, 
RAs' abstracts are equally important to writers 
and readers. 
Recent research on abstract requires manually 
analysis, which is time-consuming and labor-
intensive. Moreover, with the rapid development 
of science and technology, learners are 
increasingly engaged in self-paced learning in a 
digital environment. Our study, therefore, 
attempts to investigate ways of automatically 
analyzing the move structure of English RAs’ 
abstracts and develops an online learning system, 
CARE (Concordancer for Academic wRiting in 
English). It is expected that the automatic 
analytical tool for move structures will facilitate 
non-native speakers (NNS) or novice writers to 
be aware of appropriate move structures and 
internalize relevant knowledge to improve their 
writing. 
2 Macrostructure of Information in 
RAs 
Swales (1990) presented a simple and succinct 
picture of the organizational pattern for a RA—
the IMRD structure (Introduction, Methods, 
Results, and Discussion). Additionally Swales 
(1981, 1990) introduced the theory of genre 
analysis of a RA and a four-move scheme, which 
was later refined as the "Create a Research 
Space" (CARS) model for analyzing a RA’s 
introduction section.  
Even though Swales seemed to have 
overlooked the abstract section, in which he did 
not propose any move analysis, he himself 
plainly realized “abstracts continue to remain a 
neglected field among discourse analysts” 
(Swales, 1990, p. 181). Salager-Meyer (1992) 
also stated, “Abstracts play such a pivotal role in 
any professional reading” (p. 94). Seemingly 
researchers have perceived this view, so research 
has been expanded to concentrate on the abstract 
in recent years. 
Anthony (2003) further pointed out, “research 
has shown that the study of rhetorical 
organization or structure of texts is particularly 
useful in the technical reading and writing 
classroom” (p. 185). Therefore, he utilized 
computational means to create a system, Mover, 
which could offer move analysis to assist 
abstract writing and reading. 
3 CARE 
Our system focuses on automatically 
computational analysis of move structures (i.e. 
41
Background, Purpose, Method, Result, and 
Conclusion) in RA abstracts. In particular, we 
investigate the feasibility of using a few 
manually labeled data as seeds to train a Markov 
model and to automatically acquire move-
collocation relationships based on a large number 
of unlabeled data. These relationships are then 
used to analyze the rhetorical structure of 
abstracts. It is important that only a small 
number of manually labeled data are required 
while much of move tagging knowledge is 
learned from unlabeled data. We attempt to 
identify which rhetorical move is correspondent 
to a sentence in a given abstract by using features 
(e.g. collocations in the sentence). Our learning 
process is shown as follows: 
 
(1)Automatically collect abstracts from the Web for 
     training 
(2)Manually label each sentence in a small set of given  
     abstracts 
(3)Automatically extract collocations from all abstracts 
(4)Manually label one move for each distinct collocation 
(5)Automatically expand collocations indicative of each 
    move 
(6)Develop a hidden Markov model for move tagging 
Figure 1: Processes used to learn collocation 
classifiers 
3.1 Collecting Training Data 
In the first four processes, we collected data 
through a search engine to build the abstract 
corpus A. Three specialists in computer science 
tagged a small set of the qualified abstracts based 
on our coding scheme of moves. Meanwhile, we 
extracted the collocations (Jian et al., 2004) from 
the abstract corpus, and labeled these extracted 
collocations with the same coding scheme.  
3.2 Automatically Expanding Collocations 
for Moves 
To balance the distribution in the move-tagged 
collocation (MTC), we expand the collocation for 
certain moves in this stage. We use the one-
move-per-collocation constraint to bootstrap, 
which mainly hinges on the feature redundancy 
of the given data, a situation where there is often 
evidence to indicate that a given should be 
annotated with a certain move. That is, given one 
collocation c
i
 is tagged with move m
i,
 all 
sentences S containing collocation c
i
 will be 
tagged with m
i
 as well; meanwhile, the other 
collocations in S are thus all tagged with m
i
. For 
example: 
 
Step 1. The collocation “paper address” 
extracted from corpus A is labeled with the “P” 
move. Then we use it to label other untagged 
sentences US (e.g. Examples (1) through (2)) 
containing “paper address” as “P” in A. As a 
result, these US become tagged sentences TS 
with “P” move. 
 
  (1)This paper addresses the state explosion problem in  
       automata based ltl model checking. //P// 
  (2)This paper addresses the problem of fitting mixture  
       densities to multivariate binned and truncated data. //P// 
 
Step 2. We then look for other features (e.g. the 
collocation, “address problem”) that occur in TS 
of A to discover new evidences of a “P” move 
(e.g. Examples (3) through (4)). 
 
  (3)This paper addresses the state explosion problem in  
       automata based ltl model checking. 
  (4)This paper addresses the problem of fitting mixture   
       densities to multivariate binned and truncated data. 
 
Step 3. Subsequently, the feature “address 
problem” can be further exploited to tag 
sentences which realize the “P” move but do not 
contain the collocation “paper address”, thus 
gradually expanding the scope of the annotations 
to A. For example, in the second iteration, 
Example (5) and (6) can be automatically tagged 
as indicating the “P” move. 
 
   (5)In this paper we address the problem of query  
       answering using views for non-recursive data log    
    queries embedded in a Description Logics 
       knowledge base. //P// 
  (6)We address the problem of learning robust  
   plans for robot navigation by observing 
       particular robot behaviors. //P// 
 
From these examples ((5) and (6)), we can 
extend to another feature “we address”, which 
can be tagged as “P” move as well. The 
bootstrapping processes can be repeated until no 
new feature with high enough frequency is found 
(a sample of collocation expanded list is shown 
in Table1).  
 
Type Collocation Move Count of 
Collocation 
with m
j
 
Total of 
Collocation 
Occurrences
NV we present P 3,441 3,668 
NV we show R 1,985 2,069 
NV we propose P 1,722 1,787 
NV we describe P 1,505 1,583 
… … … … … 
Table 1: The sample of the expanded collocation 
list 
42
3.3 Building a HMM for Move Tagging 
The move sequence probability P(t
i
+1  x t
i
) is 
given as the following description: 
We are given a corpus of unlabeled abstracts A 
= {A
1
,…, A
N
}. We are also given a small labeled 
subset S = {L
1
,…, L
k
} of A, where each abstract 
L
i
 consists of a sequence of sentence and move 
{t
1
, t
2
,…, t
k
}. The moves t
i
 take out of a value 
from a set of possible move M = {m
1
,m
2
,…,m
n
}. 
Then 
1
1
(| )
(|)
()
ii
ii
i
Nt t
Pt t
Nt
+
+
⎛⎞
=
⎜⎟
⎝⎠
 
According to the bi-gram move sequence 
score (shown in Table 2), we can see move 
sequences follow a certain schematic pattern. For 
instance, the “B” move is usually directly 
followed by the “P” move or “B” move, but not 
by the “M” move. Also rarely will a “P” move 
occur before a “B” move. Furthermore, an 
abstract seldom have a move sequence wherein 
“P” move directly followed by the “R” move, 
which tends to be a bad move structure. In sum, 
the move progression generally follows the 
sequence of "B-P-M-R-C". 
 
Table 2: The score of bi-gram move sequence 
(Note that “$” denotes the beginning or the 
ending of a given abstract.) 
 
Finally, we synchronize move sequence and 
one-move-per-collocation probabilities to train a 
language model to automatically learn the 
relationship between those extracted linguistic 
features based on a large number of unlabeled 
data. Meanwhile, we set some parameters of the 
proposed model, such as, the threshold of the 
number of collocation occurring in a given 
abstract, the weight of move sequence and 
collocation and smoothing. Based on these 
parameters, we implement the Hidden Markov 
Model (HMM). The algorithm is described as the 
following: 
1111
( ,...., ) ( ) ( | ) ( | ) ( | )
niii
ps s pt ps t pt t ps t
−
= Π  
The moves t
i
 take out of a value from a set of 
possible moves M={m
1
, m
2
, …., m
k
} (The 
following parameters θ
1
 and θ
2
 will be 
determined based on some heuristics). 
(| )
ii i
p St m=  
= θ
1
 if S
i
 contains a collocation in MTC
j
 
 ij=  
= θ
2
 if S
i
 contains a collocation in MTC
j
 
 but ij≠  
= 
1
k
 if S
i
 does not contain a collocation MTC
j
 
The optimal move sequence t* is 
12
12 1
, ,...,
( *, *,..., *) ( ,..., | ,..., )arg max
n
nnin
tt t
tt t ps st t=  
In summary, at the beginning of training time, 
we use a few human move-tagged sentences as 
seed data. Then, collocation-to-move and move-
to-move probabilities are employed to build the 
HMM. This probabilistic model derived at the 
training stage will be applied at run time. 
4 Evaluation 
In terms of the training data, we retrieved 
abstracts from the search engine, Citeseer; a 
corpus of 20,306 abstracts (95,960 sentences) 
was generated. Also 106 abstracts composed of 
709 sentences were manually move-tagged by 
four informants. Meanwhile, we extracted 72,708 
collocation types and manually tagged 317 
collocations with moves.  
At run time, 115 abstracts containing 684 
sentences were prepared to be the training data. 
We then used our proposed HMM to perform 
some experimentation with the different values 
of parameters: the frequency of collocation types, 
the number of sentences with collocation in each 
abstract, move sequence score and collocation 
score.  
4.1 Performance of CARE 
We investigated how well the HMM model 
performed the task of automatic move tagging 
under different values of parameters. The 
parameters involved included the weight of 
transitional probability function, the number of 
sentences in an abstract, the minimal number of 
instance for the applicable collocations. Figure 2 
indicates the best precision of 80.54% when 627 
sentences were qualified with the set of various 
Move t
i
 Move t
i+1
 - log P (t
i+1
|t
i
) 
$ B 0.7802 
$ P 0.6131 
B B 0.9029 
B M 3.6109 
B P 0.5664 
C $ 0.0000 
M $ 4.4998 
M C 1.9349 
M M 0.7386 
M R 1.0033 
P M 0.4055 
P P 1.1431 
P R 4.2341 
R $ 0.9410 
R C 0.8232 
R R 1.7677 
43
parameters, including 0.7 as the weight of 
transitional probability function and a frequency 
threshold of 18 for a collocation to be applicable, 
and the minimally two sentences containing an 
applicable collocation. Although it is important 
to have many collocations, it is crucial that we 
set an appropriate frequency threshold of 
collocation so as not to include unreliable 
collocation and lower the precision rate. 
 
Figure2: The results of tagging performance with 
different setting of weight and threshold for 
applicable collocations (Note that C_T denotes 
the frequency threshold of collocation) 
5 System Interface 
The goal of the CARE System is to allow a 
learner to look for instances of sentences labeled 
with moves. For this purpose, the system is 
developed with three text boxes for learners to 
enter queries in English (as shown in Figure3.): 
• Single word query (i.e. directly input one 
word to query)  
• Multi-word query (i.e. enter the result 
show to find citations that contain the 
three words, “the”, “paper” and “show” 
and all the derivatives) 
• Corpus selection (i.e. learners can focus on 
a corpus in a specific domain)  
Once a query is submitted, CARE displays the 
results in returned Web pages. Each result 
consists of a sentence with its move annotation. 
The words matching the query are highlighted. 
Figure 3: The sample of searching result with the 
phrase “the result show” 
6 Conclusion 
In this paper, we have presented a method for 
computational analysis of move structures in 
RAs' abstracts and addressed its pedagogical 
applications. The method involves learning the 
inter-move relationships, and some labeling rules 
we proposed. We used a large number of 
abstracts automatically acquired from the Web 
for training, and exploited the HMM to tag 
sentences with the move of a given abstract. 
Evaluation shows that the proposed method 
outperforms previous work with higher precision. 
Using the processed result, we built a prototype 
concordance, CARE, enriched with words, 
phrases and moves. It is expected that NNS can 
benefit from such a system in learning how to 
write an abstract for a research article. 
References 
Anthony, L. and Lashkia, G. V. 2003. Mover: A 
machine learning tool to assist in the reading and 
writing of technical papers. IEEE Trans. Prof. 
Communication, 46 185-193.  
American National Standards Institute. 1979. 
American national standard for writing abstracts. 
ANSI Z39, 14-1979. New York: Author.  
Jian, J. Y., Chang, Y. C., and Chang, J. S. 2004. 
TANGO: Bilingual Collocational Concordancer, 
Post & demo in ACL 2004, Barcelona. 
Salager-Meyer, F. S. 1992. A text-type and move 
analysis study of verb tense and modality 
distribution in medical English abstracts. English 
for Specific Purposes, 11:93-113.  
Swales, J.M. 1981. Aspects of article introductions. 
Birmingham, UK: The University of Aston, 
Language Studies Unit.   
Swales, J.M. 1990. Genre analysis: English in 
Academic and Research Settings. Cambridge 
University Press   
44
