IEEE Consumer Electronics Magazine - November/December 2021 - 91

crawling or sensing module, which crawls socialmedia
information from different sources, such
as Twitter, Telegram, and Really Simple Syndication
(RSS) messages published by Spanish media,
2) the NLP processing module, which analyzes
and categorizes these microtexts into different
topics previously defined by the user, setting a
topic score, sentiment score, and combines them
into a PPI score, and 3) the visualization module,
which shows the processed, geolocated, and
categorized information, so that it can provide
insights for the user. Moreover, the raw and processed
information is stored in an Elastic Search
database. In the remainder of this section, we
introduce the design and implementation details
of WATERSensing's main modules.
Crawling Module
The crawling module is responsible for
obtaining a collection of microtexts from different
social networks. Particularly, we have
focused on Twitter, Telegram, and RSS, based on
user-defined settings such as a list of RSS feeds,
URLs, Twitter hashtags and/or keywords to be
tracked, and Telegram channels/groups. The
acquisition of tweets is performed through the
Twitter API by setting specific keywords (e.g.,
aquifer, drought, landslide,or waste dumping). A
fixed list of RSS feeds is established to track
information from representative Spanish media
sources. These include national and regional
newspapers, official websites of public institutions
and government bodies, etc. Finally, the
Telegram crawler is linked to the user account of
the admin user, filtering private chats. Duplicate
feeds are filtered by checking an MD5 hash generated
for each microtext, which was previously
stored in the storage module.
NLP Processing Module
The NLP processing module of the application
is based on CASPER,13 a workbench that
is aimed at detecting multidomain problems
described in microtexts. The main tasks in this
module are as follows.
Setting Topic Categories One or more categories
of user-defined topics (e.g., drought, flooding,
etc.) must be previously registered. This
category is called " space " in our system, and it is
November/December 2021
created semiautomatically by selecting significant
features, which are relevant words related
to the topic of interest. We use seven WordNet
semantic relationships to expand the initial keywords,
i.e., x-causes-y, x-derived_from-y, x-has_
hyponym-y, x-has_subevent-y, x-near_synonym-y,
x-pertains_to-y, and x-related_to-y.
The first step is to select some seed terms
that are representative of each category. Currently,
the semantic expansion is based on Spanish.
In this example, the keyword is " lluvia " (i.e.,
rain), which is the same word used to name the
space (in capital letters). Second, the user must
also select the relevant meanings to which each
keyword is linked, so that the system scans the
glosses for definitions related to the given keyword.
In this example, the second meaning of
" lluvia " corresponds to " Atmospheric phenomenon
consisting of rain, water falling from the
clouds, from the sky, " which is represented by
the term precipitacon (i.e., rainfall) and associated
to words such as aguacero, precipitacion,
lluvia, and pluvioso (i.e., downpour, rainfall, rain,
pluvial, etc.). Then, each selected keyword is
converted into a topical feature, resulting in vector
Ci ¼ðfi1;fi2; ... ;fikÞ, where every fij identify
a feature in the form of a WordNet synset (i.e.,
set of synonymous words) assigned to the category
Ci. Third, the system suggests a set of additional
words and phrases from these synsets, so
that the user can choose them to augment the
original list of category descriptors. In this example,
some additional words related to the first
meaning of " lluvia " are agua dulce and bajada
descenso (i.e., fresh water and descent).
Processing Natural Language The main
goal of this stage is to perform the aggregation
and filtering of the monitored topics. Thus,
each microtext is divided into sentences,
each sentence is tokenized, and then each token
is labeled with a part of speech (POS). A tweet
is now represented as the vector Tm ¼
ðwm1;wm2; ...;wmpÞ, in which wmn is an object for
every word that appears in the tweet Tm, being p
the number of words of Tm. Each wmn is represented
with several attributes, such as the word
form, the position in the microtext, the stem,
the POS, the topic, and the sentiment, just to
mention a few.
91
IEEE Consumer Electronics Magazine - November/December 2021

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