O n f e r e n c e

 
Generating RWT Feedback
The RWT web-based system that we developed allows users to input Introduction, Methods, 
Results and Discussion sections of their research articles so that each sentence can be 
classified as belonging to a specific move and a specific step within that move. In addition 
to providing sentence-
level feedback, the RWT compares a user’s article to corpus articles in 
the same discipline. More specifically, the RWT interface informs the user whether or not a 
given step is over-represented or under-represented in their writing sample compared to 
published articles in that same discipline.
 
Conclusion and Future Work 
This paper described an approach to automated writing evaluation of RA Introduction 
sections using sentence-level rhetorical function analysis. This approach is demonstrative of 
the analysis framework that consists of a set of machine learning classifiers trained to 
rec
ognize a sentence’s move and step for all RA sections (Introduction, Methods, Results, 
and Discussion). The classifiers use the 
k
-gram feature representation for text documents 
supplemented by both stemmed text and the corresponding part-of-speech tags. The 
machine learning algorithm is implemented using the Maximum Entropy method for 
document classification. The implementation has been integrated into the Research Writing 
Tutor to provide sentence-level and section-
level feedback on students’ research arti
cles. 
Future work will focus on improving the performance of the classifiers in order to provide 
more accurate and more useful feedback to RWT users. Preliminary results show that both 
move and step recognition rates can be improved substantially by using additional features 
in the form of hand-crafted regular expressions that capture common patterns associated 
with different rhetorical functions.
In addition to improving classification rates, we plan to conduct several usability studies to 
assess the effectiveness of the classification system as well as the web-based interface used 
by the RWT website.
 
0
0,1
0,2
0,3
0,4
0,5
0,6
Introduction
Methods
Results
Discussion
Cl
ass
if
icati
o
n
Perf
o
rm
an
ce
(
ka
p
p
a
) 
Move
Step

---

-122- 
2014 CALL Conference 
LINGUAPOLIS
www.antwerpcall.be