Page 138 - Handbook of Deep Learning in Biomedical Engineering Techniques and Applications

Chapter 5 Depression discovery in cancer communities using deep learning 127

Authors in Ref. [28] detect favorable and unfavorable opinions
from web pages and news articles using a sentiment lexicon. They
first identify the subjective or sentiment bearing expressions and
then determine the relationship of these expressions to their sub-
ject. Finally, they compute document sentiment by evaluating the
polarity of the subjective expressions. They ticket verbs as senti-
ment transmitters that transfer sentiment from one argument to
the other. For example, in the sentence “this book is good,” the
verb is transmits the sentiment extracted from good to the subject
book and associates them. Terms having part-of-speech tags
different from verb are dealt with in an easier waydthey directly
transfer their sentiment to the related argument. There are certain
weaknesses of their approach. They are unable to resolve corefer-
ences or expressions that refer to the same entity in text. More-
over, they do not have in place any mechanism for dealing with
negations.

2.2Machine learningebased approaches
ML approaches incorporate the famous ML algorithms that
treat SA as a regular text classification problem that makes use
of linguistic features. Most ML-based methods use a supervised
ML classifier.

2.2.1 Supervised machine learning
Probabilistic or generative classifiers: This set of classifiers use
mixture models for grouping and work on the assumption that
each of the categories is a component of the mixture. Each
mixture constituent is a generative model that offers the likeli-
hood of selection of a specific word for that constituent. The three
most famous generative classifiers are Bayesian network (BN),
naïve Bayes (NB) classifier, and maximum entropy classifiers
(MaxEnt).
Linear classifiers: Given ¼ {x1, . xn} is the normalized docu-
ment word frequency, vector ¼ {a1, . an} is a vector of linear co-
efficients with the same dimensionality as the feature space, and b
is a scalar; the output of the linear predictor is defined as p ¼þb,
which is the output of the linear classifier. The predictor p is a
separating hyperplane between different classes. Two of the
most famous linear classifiers are ṣupport vector machines
(SVMs) [29], [30] and neural networks (NNs).
Decision tree classifiers: Decision tree classifier offers a tiered
disintegration of the training data space. In this, a condition
on the feature value splits the data [31]. The condition is the