Digital Signal Processing Reference
In-Depth Information
Table 10.12
WA f o r
different BoNG feature types
Feature type
WA [%]
f
i
,
j
75.03
Norm(
f
i
,
j
)
75.72
TF
75.44
Norm(TF)
75.66
IDF
75.05
Norm(IDF)
76.42
TFIDF
75.45
Norm(TFIDF)
77.14
Table 10.13
WA during
N-gram length optimisation
from
g
min
to
g
max
for BoNG
features
g
min
g
max
# Features
WA [%]
1
1
18 316
74.58
1
2
96 152
75.95
1
3
151 083
77.14
1
4
171 438
76.41
1
5
177 733
76.92
2
2
77 840
66.72
2
3
132 780
66.54
2
4
153 146
69.62
2
5
159 465
71.66
3
3
54 968
71.59
3
4
75 418
72.33
3
5
81 911
72.61
and normalisation methods were evaluated. Table
10.12
summarises the obtianed
WAs for simple N-Gram frequency (
f
i
,
j
), TF, IDF, and normalisation (norm) for
N-Grams from one to three terms (i.e.,
g
min
=
1 and
g
max
=
3). It will be shown
that this is an optimal choice.
Little difference is observed for the different types of feature representation, of
which normalisation combined with TFIDF leads to the best result, and normalisation
improves results at any time and in particular in the case of IDF.
Let us now consider the optimal N-Gram length from
g
min
to
g
max
in Table
10.13
.
As stated, the optimal choice is
g
min
=
3. This agrees with [
120
], where
optimal results for product reviews where reached by tri-grams. Yet, the authors of
this study found back-off N-Grams to downgrade the accuracies—this is different to
the case of Metacritic. In this optimal setting, 12 % of the features are single words,
52 % bi-grams, and 36 % tri-grams. The largest feature space in this evaluation has
177 733 features for
g
min
=
1 and
g
max
=
1 and
g
max
=
5, the smallest one only 18 316 features
for
g
min
=
1.
Looking next at out-of-vocabulary words (OOV), i.e., such occurring in the test,
but not in the training material, these make up 30.3 % of the total vocabulary. Out-
of-vocabulary events, i.e., the number of occurrences of OOV words in all reviews,
1 and
g
max
=
