IEEE Consumer Electronics Magazine - November/December 2021 - 92

Special Section on Advances in Smart Cities
The topic value of each wmn is obtained as follows.
The ngrams (i.e., sequences of n contiguous
lexical items in a given text) of each wmn in
Tm are searched in the vector Ci. If it matches
any fij in Ci, the topic value is set to 1. Calculating
the value of sentiment involves detecting significant
ngrams with respect to sentiment. These
significant ngrams are discovered by making use
of two distinct datasets. On the one hand, a
polarity lexicon called SENTIMENTS holds the
polarity type (i.e., positive or negative) for a
number of words, where their POS (i.e., verb,
adjective, noun, or adverb) is taken into consideration.
On the other hand, the user can set any
fij in Ci as a negative word when setting up topic
categories. This feature allows users to expand
the lexicon of polarity with additional negative
words that are related to the particular context
domain. Thus, the ngrams of each wmn in Tm are
searched in both datasets. The application
assigns the values +1 or -1 (for positively and
negatively marked ngrams, respectively) to the
feeling attribute of each wmn in Tm when a hit is
found.
Determining the Topic Score Since tweets
and categories are represented as vectors, a
measure of similarity can be used to evaluate
the degree of relationship between these two
vectors. In this context, the " cosine similarity "
(or normalized point product) is used as a measure
of semantic distance.
Determining the Sentiment Score Our
application does not use a straightforward
approach to sentiment calculation, i.e., just summing
up the sentiment values of the ngrams in
the microtext and determining the polarity of
the message by the sign of the final score. Our
sentiment measure is grounded on the degree of
sentiment in each tweet through a metric originally
used to assess political positions in texts.14
In particular, a scaling procedure changes the
counts of sentiment-coded ngrams in the tweet
Tm into a point on the sentiment dimension S.It
is important to note that, as we are only concerned
with problem detection, the sentimentrelatedness
function automatically assigns the
value of zero to those tweets that do not express
negative polarity.
92
Detecting the Problem With the aim of having
a final score that could serve to identify if a
particular tweet is likely to deal with a given
water-related problem, the PPI is eventually
computed, resulting from the geometric mean of
the topic score and the sentiment score. Therefore,
tweets can be sorted according to the PPI,
which can be used to establish alert thresholds,
so that messages can be sent to the system
administrator to report about any potential risk,
hazard, or disaster. This alert system is now
being implemented through e-mail notifications.
Actually, when there are more than N messages
with PPI > 0:5, the system sends an e-mail to
the administrator with a summary of these
messages.
Detecting Entities We are also capable of
labeling up to three different types of entities,
i.e., LOC for location, PER for person, and ORG
for organization, through a named-entity recognition
(NER) module. This module is built using
the Spacy library, which has a pretrained model
for Spanish. However, for a better performance,
a custom model has been trained using a corpus
automatically extracted from Wikipedia. The
framework uses a transition-based prediction
model (i.e., embedding plus encoding with a convolutional
neural network) capable of understanding
the context of each word in a sentence,
and using this context as input for the neural network
to predict its next move, i.e., whether it
tags a word as a named entity or as part of a
named entity, skip to the next word, etc.
Visualizing Results
The visualization module shows the information
collected and processed by the previous
modules. We refer the reader to sensingtools.com
and click WATERSensing for insights. It shows
the number of feeds that have been categorized
from different sources (e.g., Twitter, RSS, and
Telegram), along with the processed messages
and the metainformation obtained with the NLP
module to identify the probability that this text
represents a problem (i.e., PPI), the main entities
recognized in the text, and the date when it was
published within the selected time frame (i.e.,
1 h, 1 day, 1 week, or 1 month). Moreover, the
processed messages can not only be filtered by
IEEE Consumer Electronics Magazine
http://sensingtools.com
IEEE Consumer Electronics Magazine - November/December 2021

Table of Contents for the Digital Edition of IEEE Consumer Electronics Magazine - November/December 2021
