Computational Intelligence - November 2015 - 29

semantic clues, some existing related works relied on using
knowledge bases [26], [27]. The bag-of-concepts (BoC) model
leverages on representing text as a conceptual vector rather than
relying on the terms in the text. For example, if a text contains
"red" and "orange," then BoC models them as the concept
"color," i.e., BoC looks for hyponym. The BoC model was first
proposed by Sahlgren et al. [28] to enhance the performance of
support vector machine (SVM) in text categorization tasks.
According to their method, concepts are the synonym sets of
BoW. Among recent approaches on the BoC model, the
approach by Wang et al. [29] presented the idea of concept as a
set of entities in a given domain, i.e., words belonging to similar
classes have similar representation. If a text contains "Jeep" and
"Honda", this can be conceptualized by the concept "car". On
the basis of their study, we identify two major advantages of the
BoC model:
❏ Replacement of surface matching with semantic similarity: the
BoC model calculates semantic similarity between words
and multi-word expressions from a higher concept level.
❏ Tolerance to new terms: In text, new terms are always encountered, but new concepts may not always arise. Once BoC
models concepts for a category, it can handle new words
under that category. This shows the strong adaptability of
the BoC model with respect to word changes.
Zhang et al. [30] discussed semantic classification on a disease
corpus. Though their approach does not focus on the BoC
model, they attempted to capture semantic information from
text at the highest level. According to them, the use of contextual
semantic features along with the BoW model can be very useful
for semantic text classification. Wu et al. [31] built a sentiment
lexicon using a common-sense knowledge base, ConceptNet.
Using the hypothesis that concepts pass their sentiment intensity
to their neighbors based on the relations connecting them, they
constructed an enriched sentiment lexicon able to produce better performance in the sentiment polarity classification task.

feel_happy. It was built using semantic multidimensional
scaling [33].The knowledge base is available in the form of RDF
XML files, as well as through a web-based API2.
B. Sentiment Flow Algorithm: An Informal Description

Consider the sentence
(1) The car is very old but it is rather not expensive.

The proposed polarity detection algorithm retrieves the polarity scores of concepts from SenticNet [32]. SenticNet is a sentiment lexicon, freely available as a knowledge base that contains
polarity scores of individual words and multi-word expressions.
The version of SenticNet used was 3.01. This common-sense
knowledge base contains 30 thousand affective concepts, such as
celebrate_special_occasion, make_mistake, and

A baseline approach to polarity detection consists in counting
positive and negative words in the sentence, using a polarity
lexicon. In this example, as it is typical for any text, most of the
words, such as "car" or "very", do not have any intrinsic polarity. Suppose the polarity lexicon contains two words: "old" and
"expensive", both with negative polarity, at least in the context
of buying a car. Thus, the total, or average, polarity of the words
in this sentence is negative.
The main idea behind the use of sentic patterns [3] for calculating the overall polarity of a sentence can be best illustrated by
analogy with an electronic circuit, in which few "elements" are
"sources" of the charge or signal, while many elements operate on
the signal by transforming it or combining different signals. This
implements a rudimentary type of semantic processing, where the
"meaning" of a sentence is reduced to only one value-its polarity.
Sentic patterns are applied to the dependency syntactic tree
of the sentence, such as the one shown in Figure 1(a). The only
two words that have intrinsic polarity are shown in yellow color;
the words that modify the meaning of other words in the manner similar to contextual valence shifters [34] are shown in blue.
A baseline that completely ignores the structure, as well as words
that have no intrinsic polarity, is shown in Figure 1(b): the only
two words in this sentence are negative, so the total is negative.
However, the syntactic tree can be re-interpreted in the form of
a "circuit" where the "signal" flows from one element, or subtree, to another, as shown in Figure 1(b): the word "combines"
two parts of the sentence, each one, in this case having a polarity
words combined with modifiers. Removing the words not used
for calculation (shown by white rectangles) and re-structuring
the tree, a sequence of transformations shown in Figure 1(d) is
obtained. In blue color shown are elements resembling electronic
amplifier, electronic logical negation, a resistor, and a kind of an
asymmetric logical "and".
Figure 1(e) illustrates this idea at work: the sentiment flow
from polarity words through shifters and combining words. The
two polarity-bearing words in this example are negative. The
negative effect of the word "old" is amplified by the intensifier
"very". However, the negative effect of the word "expensive" is
inverted by the negation, and the resulting positive value is
decreased by the "resistor". Finally, the values of the two phrases
are combined by the word "but", so that the overall polarity still
has the same sign as that of the second component, but further
decreased in value by the first component. Note that the effect of
the conjunction would be opposite for, for example, "although":

III. Sentic Pattern Rules

Sentic pattern rules are a recently suggested semantic technique
for advanced analysis of text with the aim of determining its
polarity [3]. In this section, we introduce the main idea of the
sentiment flow algorithm based on such rules and give examples of novel rules, which are used in conjunction with the
ones previously described in [3]. In the next section, the sentiment flow algorithm will be described in more detail.
A. Resource Used: SenticNet

http://sentic.net/senticnet-3.0.zip

http://sentic.net/api/

November 2015 | Ieee ComputatIoNal INtellIgeNCe magazINe

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