Digital Signal Processing Reference
In-Depth Information
deception can be employed to adapt the system and the teaching pace [
109
,
110
];
generally, paralinguistic cues are essential for tutors and students to make learn-
ing successful [
111
]. In addition, automatic voice coaching, e.g.,to give better pub-
lic speeches or simply to intonate appropriately when learning foreign languages,
becomes possible [
112
].
Moreover, there are many security related situations (surgical operation [
113
],
crisis management, and all the tasks connected to aviation and air traffic control)
where monitoring of stress level, sleepiness, intoxication, and such, may play a vital
role [
114
]. In addition, counter terrorism or counter vandalism surveillance may be
aided by analysing paralinguistic cues such as aggressiveness of potential aggressors
[
115
,
116
], or fear of potential victims [
117
].
Finally, in the field of multimedia retrieval, paralinguistic information is of interest
for manifold types of media searches [
4
,
118
].
10.4.1 Sentiment and Opinion
Sentiment analysis and opinion mining have been studied for manifold application
scenarios such as product reviews [
119
-
124
], the stock market [
125
], or hotels and
travel destinations [
119
,
122
], and film reviews [
119
,
126
,
127
]. A particularly
difficult task among these is sentiment analysis for film reviews: In [
119
], a 66 %
WA for valence polarity classification are named for film reviews, but 84 % WA for
car reviews by identical means of analysis. This may be owed to the discrepancy
between the semantic orientation of words describing the elements of a film (i.e., a
scene, the plot), and its style or art.
In this light this section introduces valence estimation from text by means of
on-line knowledge sources and machine learning. The experiments are reported on a
database of over 100 k film reviews collected from the review website Metacritic
4
—
the largest to-date. Metacritic's fine-grained review scores as gold standard allow for
a regression approach besides binary or ternary valence prediction.
10.4.1.1 Metacritic Database
The database described in this section is likely the largest film review corpus to
date. Manifold other do, however, exist, ranging in size from 120 to 2 000 reviews.
In [
119
], 120 film reviews were collected from Epinions.
5
In [
127
]11filmswere
selected from the top 250 list of IMDB.
6
Then, the first 100 reviews were retrieved
for each one of them resulting in a total of 16 k sentences and 260 k words. Perhaps
the most frequently used film review database was introduced in [
126
]. In the begin-
4
5
6
