Digital Signal Processing Reference
In-Depth Information
led to significant gains. The results based on a three-class task, but alternatively
regression by SVR was evaluated.
In the future, more specific term categories could be used from General Inquirer.
Further, in [
138
] it was shown that ConceptNet can also be exploited directly for
sentiment information. Thus, it could complement General Inquirer in this respect.
Also, multi-word terms and complete phrases could be used directly in ConceptNet.
Of particular help could also be the addition of a named entity recognition in com-
bination with other types of common knowledge sources such as Wikipedia. Next,
out-of-vocabulary resolution could be improved by using more WordNet relations,
such as
antonymy
, i.e., opposite meaning, or
hypernymy
, i.e., more general mean-
ing. Finally, OOV N-Grams need to be resolved by a second substitution step after
N-Gram creation.
10.4.2 Short-term States: Emotion and Interest
The recognition of a number of short-term states from speech has been addressed so
far, of which the following non-exhaustive list names some examples:
•
mode
: speaking style [
139
] and voice quality [
140
];
•
emotions
(full-blown, prototypical): [
141
];
•
emotion-related states or affects
: for example, general [
142
-
144
], stress [
145
],
intimacy [
146
], interest [
65
,
75
], confidence [
147
], uncertainty [
107
,
148
], decep-
tion [
149
,
150
], politeness [
151
-
153
], frustration [
154
-
156
], sarcasm [
157
,
158
],
pain [
99
].
From these, two examples among most researched candidates have been cho-
sen for illustration of methodology and performances: emotion and interest, both
belonging to 'affective' speaker states.
The young field of affect recognition from voice has recently gained considerable
interest in the fields of Human-Machine Communication, Human-Robot Commu-
nication, and Multimedia Retrieval. Numerous studies have been seen in the last
decade trying to improve on features and classifiers [
159
]. One first cooperative
experiment is found in the CEICES initiative [
160
], where seven sites compared
their classification results under exactly the same conditions and pooled their fea-
tures together for one combined unified selection process. This comparison was not
fully open to the public, which was the motivation to create the INTERSPEECH 2009
Emotion Challenge—the first in an ongoing series of challenges on Computational
Paralinguistics—which are conducted for strict comparability: all participants use
the same database and the same evaluation measures in their experiments. As classes
are unbalanced, the primary measure to optimise is UA (unweighted average recall),
and secondly WA (weighted average recall by number of instances per class—this
is commonly known simply as “accuracy”).
