ORDSAENet: Outlier Resilient Semantic Featured Deep Driven Sentiment Analysis Model for Education Domain

Abstract

The high pace rising global competitions across education sector has forced institutions to enhance aforesaid aspects, which require assessing students or related stakeholders’ perception and opinion towards the learning materials, courses, learning methods or pedagogies, etc. To achieve it, the use of reviews by students can of paramount significance; yet, annotating student’s opinion over huge heterogenous and unstructured data remains a tedious task. Though, the artificial intelligence (AI) and natural language processing (NLP) techniques can play decisive role; yet the conventional unsupervised lexicon, corpus-based solutions, and machine learning and/or deep driven approaches are found limited due to the different issues like class-imbalance, lack of contextual details, lack of long-term dependency, convergence, local minima etc. The aforesaid challenges can be severe over large inputs in Big Data ecosystems. In this reference, this paper proposed an outlier resilient semantic featuring deep driven sentiment analysis model (ORDSAENet) for educational domain sentiment annotations. To address data heterogeneity and unstructured-ness over unpredictable digital media, the ORDSAENet applies varied pre-processing methods including missing value removal, Unicode normalization, Emoji and Website link removal, removal of the words with numeric values, punctuations removal, lower case conversion, stop-word removal, lemmatization, and tokenization. Moreover, it applies a text size-constrained criteria to remove outlier texts from the input and hence improve ROI-specific learning for accurate annotation. The tokenized data was processed for Word2Vec assisted continuous bag-of-words (CBOW) semantic embedding followed by synthetic minority over-sampling with edited nearest neighbor (SMOTE-ENN) resampling. The resampled embedding matrix was then processed for Bi-LSTM feature extraction and learning that retains both local as well as contextual features to achieve efficient learning and classification. Executing ORDSAENet model over educational review dataset encompassing both qualitative reviews as well as quantitative ratings for the online courses, revealed that the proposed approach achieves average sentiment annotation accuracy, precision, recall, and F-Measure of 95.87%, 95.26%, 95.06% and 95.15%, respectively, which is higher than the LSTM driven standalone feature learning solutions and other state-of-arts. The overall simulation results and allied inferences confirm robustness of the ORDSAENet model towards real-time educational sentiment annotation solution.