Abstract:
"When it comes to recommendation systems which rank products through sentiment analysis,
the sentiment scores of emojis are not considered. Also, emphasized informal words such as
‘liiiike’, ‘puuuuurfct’, or ‘yasssss’ are very common in user generated text contents in various
online platforms. However, these words cannot be preprocessed properly with existing
methods and will lead to poor correctness on tasks such as sentiment analysis which is a part
of the NLP (Natural Language Processing) domain. These words are also not considered in
sentiment analysis based recommendation systems which will affect in poor correctness on
product rankings.
With the use of machine learning, the problem of emphasized informal words has been
addressed by training a model based on a supervised classification algorithm with a
synthesized dataset of emphasized words so that this model can predict the actual word which
is meant by the emphasized counterpart. With this model as well as another model which
considers emojis when analyzing sentiment, the correctness of sentiment based ranking
process has been improved.
Tests on emphasized informal words model yielded good results on standard evaluation
metrics. It also performs well on words with intentional misspellings with repeated letters
while the ranking process was also improved according to the benchmarks."
