Abstract:
Millions of tweets are posted on Twitter related to different events and situations. As a result, many research problems have been identified for Twitter
data. One of them is tweet classification, especially classification of tweets
posted during critical situations, disasters, or in the time of need. Request
classification is a sensitive issue as it helps to save precious lives. In previously proposed methodologies, a lot of feature extraction techniques have
been used for tweet classification. These techniques lack a lot of aspects that
should be considered for request classification. Features extracted by those
techniques do not contain any contextual information of the text. Moreover,
features are limited to word level. Another issue is the curse of dimensionality due to n-gram features. Metadata is also used as features in some of the
proposed methodologies and it did not contribute to the performance of the
framework. Improvement in feature extraction methods can lead to better
results. This work aims to develop a deep learning-based framework which
extracts contextual features from tweets. These contextual features are more
reliable, avoid the curse of dimensionality and capture semantic information.
Moreover, sentence-level feature has also been extracted from tweets. Three
different feature sets have been extracted to achieve maximum output. Multiple classifiers have been used to validate the performance of this framework.
In the end, 612-dimensional hybrid features have been created and experimented using four different classifiers. Neural Networks outperformed other
classifiers and produced improved results as compared to the baseline model.
The proposed framework achieved 89% accuracy, 96% precision, 90% recall,
and 93% F1- score using hybrid features with Hold-out validation.
