Computational Intelligence - November 2015 - 35

accuracy of 88.27% at the sentence level. We tested the performance of the other benchmark sentiment analysis systems on
this dataset. As on Movie Review dataset, the new patterns and
new training sets increased the accuracy over [3]. Further, the
method by Socher et al. [7] was found to perform very poorly
on the Blitzer dataset.
3) Results on the Amazon Dataset: The same table shows the
results of dynamic sentic patterns on the Amazon dataset
described in Section V-A3. Again, the proposed method outperforms the state-of-the-art approaches.
B. Discussion

The proposed approach outperforms the state-of-the-art methods on both the movie review and the Amazon datasets. The
results achieved on the Blitzer-derived dataset are even more
impressive. This can be considered as evidence of robustness of
the proposed method, since its high performance is stable
across very different datasets in different domains. Moreover,
while standard statistical methods require extensive training,
both in terms of resources (training corpora) and time (learning time), sentic patterns are mostly unsupervised, except for
the dynamic CI module, which is, though, very fast, due to the
use of ELM.
The addition and improvement of the patterns, as noted in
[3], has helped the system improve its results. Results show performance improvement over [3]. On the other hand, [7] has
failed to obtain consistently good accuracy over both Blitzer
and amazon datasets but obtained good accuracy over the
Movie Review dataset. This is because the classifier proposed in
[7] was trained on the Movie Review dataset only.
The proposed approach has therefore obtained a better
accuracy than the baseline system. We combined the three
datasets described in Section V-A1, V-A2 and V-A3 to evaluate
sentic patterns. From Section V-A, we can calculate the number
of positive and negative sentences in the dataset, which shows
72,721 positive and 56,903 negative sentences. If the system
predicts all sentences as positive, this would give a baseline
accuracy of 56.10%. Clearly, the proposed system significantly
outperformed the baseline system. Since the performance of
the proposed method relies on the quality of dependency parsing, which in turn requires grammatically correct text,
ungrammatical sentences present in all three datasets negatively
affected results. It is worth noting that the accuracy of the system crucially depends on the quality of the output of the
dependency parser, which relies on grammatical correctness of
the input sentences.
On the other hand, compilation of a balanced training dataset has a strong impact on developing a more accurate classifier
than the one reported by Poria et al. [3].
1) Results obtained using SentiWordNet: We carried out an
extensive experiment using SentiWordNet instead of SenticNet
on all the three datasets. The results showed SenticNet performed slightly better than SentiWordNet. A possible future
direction of this work is the invention of a novel approach to
combine SenticNet and SentiWordNet in the sentiment analysis

Table 4 Precision achieved with different classification
algorithms on different datasets.
bliTzermovie review derived
daTaseT
daTaseT
80.00%
-

amazon
daTaseT
-

RNtN (SocheR et al.
2013 [7])
PoRia et al. 2014 [3]

85.40%

61.93%

68.21%

86.21%

87.00%

79.33%

SeNtic PatteRNS

87.15%

86.46%

80.62%

elM claSSifieR

71.11%

74.49%

71.29%

eNSeMble claSSificatioN 88.12%
(PRoPoSed Method)

88.27%

82.75%

algoriThm
RNN (SocheR et al.
2012 [38])

Table 5 Results obtained using SentiWordNet.
daTaseT

Using senTicneT

Using senTiwordneT

Movie Review

88.12%

87.63%

blitzeR

88.27%

88.09%

aMazoN

82.75%

80.28%

Table 6 Some examples where a CI classifier was used to
obtain the polarity label.
senTence

PolariTy

i had to RetuRN the PhoNe afteR 2 dayS of uSe.

Negative

the PhoNe RuNS ReceNt oPeRatiNg SySteM.

PoSitive

the PhoNe haS a big aNd caPacitive touchScReeN.

PoSitive

My iPhoNe batteRy laStS oNly few houRS.

Negative

i ReMeMbeR that i SlePt at the Movie hall.

Negative

framework. The slight difference in the accuracy reported in
Table 5 confirmed that both the lexicons share similar knowledge but since SenticNet contains concepts, this helps increase
accuracy. For example, in the sentence "The battery lasts little",
proposed algorithm extracts the concept "last little" which exists
in SenticNet but not in SentiWordNet. As a result, when SenticNet is used the framework labels the sentence with a "negative"
sentiment but when using SentiWordNet the sentence is labeled
with a "neutral" sentiment.
2) Examples of cases when a CI classifier was used: Table 6 presents some examples where the CI module was used to guess
the polarity. For each of these sentences, no concept was found
in SenticNet.
VII. Conclusions

The paper shows how computational intelligence and linguistics can be blended in order to understand sentiment associated
with the text. The presented approach combines the use of various linguistic patterns based on the syntactic structure of the
sentences. Similarly to the function of electronic logic gates, the
algorithm determines the polarity of each word and flows, or
extends, it through the dependency arcs in order to determine
the final polarity label of the sentence. Overall, the proposed

November 2015 | Ieee ComputatIoNal INtellIgeNCe magazINe

Table of Contents for the Digital Edition of Computational Intelligence - November 2015