An Emotional Compass: Emotions on Social Networks and a New Experience of Cities - Augmented Reality Art - page 235

In the application we have, thus, chosen to gather geo-location information
through explicit use of the location-based features of the services and, should they
not have been provided, by combining them the results of the named places analysis.

14.4.3 Natural Language Processing and Artificial

Intelligence to Recognize Emotions and Topics in Text

There is an extensive amount of research about the possibility to automatically
interpret text to understand the emotion expressed by the writer, either on social
networks or on more general texts.
We approached the possibility to recognize emotions by identifying in text the co-
occurrence of words or symbols that have explicit affective meaning. As suggested
in by Ortony et al. ( 1987 ) we must separate the ways in which we handle words that
directly refer to emotional states (e.g.: fear, joy) from the ones which only indirectly
reference them, based on the context (e.g.: “killer” can refer to an assassin or to a
“killer application”): each has different ways and metrics for evaluation.
For this, we have used the classification found in the WordNet (Fellbaum 1998 )
extension called WordNet Affect (Strapparava and Valitutti 2004 ).
The approach we used was based on the implementation of a variation of the
Latent Semantic Analysis (LSA). LSA yields a vector space model that allows for a
homogeneous representation (and hence comparison) of words, word sets, sentences
and texts. According to Berry ( 1992 ), each document can be represented in the LSA
space by summing up the normalized LSA vectors of all the terms contained in it.
Thus a synset in WordNet (and even all the words labeled with a particular emotion)
can also be represented in this way. In this space an emotion can be represented at
least in three ways: (i) the vector of the specific word denoting the emotion (e.g.
“anger”), (ii) the vector representing the synset of the emotion (e.g. f anger, choler,
ire g ), and (iii) the vector of all the words in the synsets labeled with the emotion.
This procedure is well-documented and used, for example in the way shown in
Strapparava and Mihalcea ( 2008 ), which we adopted for the details of the technique.
We adapted the technique found in Iaconesi and Persico ( 2012 ) to handle multiple
languages by using the meta-data provided by social networks to understand in
which language messages were written in and using a mixture of the widely
available WordNet translations and some which we produced during the research
for specific use cases.
An annotation system was created on the databases to tag texts with the relevant
emotions (as, within the same message, multiple emotions can be expressed). For
example, Fig. 14.4 shows the results of a full week of emotional harvesting in the
city of Rome, for the emotion “trust”.
We also tried to deal with the wide presence of irony, jokes and other forms of
literary expression which are difficult to interpret automatically. To do this, we have
followed the suggestions described in Carvalho et al. ( 2009 ) and in Bermingham
and Smeaton ( 2010 ) with varying results.