Please enable JavaScript.
Coggle requires JavaScript to display documents.
SOFTCARDINALITY: Hierarchical Text Overlap for Student Response Analysis
…
SOFTCARDINALITY: Hierarchical Text Overlap for Student Response Analysis
Jimmenez et al., 2013.
Model (supervised)
-
threeclassification models (2 way, 3way and 5 way) were
earned from the training partitions on each vector
data set using a J48 graft tree (Webb, 1999)
All 6 resulting classification models were boosted with
15 iterations of bagging (Breiman, 1996). The used
implementation of this classifier was that included
in WEKA v.3.6.9 (Hall et al., 2009). => C4.5 decision tree
An Improved System:
Firstly, we decided to include in our feature set the 8 features of the lexical overlap baseline described by Dzikovska et al. (2012)...
Features
extended feature set form soft cardinality measures (table 2 on p. 282.), for Q, A and RA
basic feature set from soft cardinality measures (table 1 on p. 202), for Q, A and RA
Second, each stemmed word t was represented in
q-grams: t[3:4] for Beetle and t[4] for SciEntsBank.
These representations obtained the best accuracies
in the training data sets.
Results
At the time when the official results were released, we observed that our submitted system performed pretty well in SciEntsBank but poorly in Beetle. Moreover, the lexical-overlap baseline outperformed our system in Beetle.
-
-
2nd place in 5-way classification: ACC = 0.513 (all - UA, UQ, UD)
Improved system: 0.730, 0.634, 0.530 = ACC (all) for improved system (unofficial results) => 1st place in all classification tasks
Particularly, our system was the best
in the largest and more challenging test set, namely
“unseen domains”.
-