Machine learning algorithms have become very frequently used in natural language processing, notably sentiment analysis, which helps determine the general feeling carried within a text. Among these algorithms, Support Vector Machines have proven powerful classifiers especially in such a task, when their performance is assessed through accuracy score and f1-score. However, they remain slow in terms of training, thus making exhaustive grid-search experimentations very time-consuming. In this paper, we present an observed pattern in SVM's accuracy, and f1-score approximated with a Lagrange polynomial.

Massive open online courses (MOOCs) are a variety of courses offered through the online mode, paid or unpaid and has evolved as an excellent learning resource for students. The structure of the course design is mainly linear where there are a few video lectures provided by either professors of several universities, or people with expertise in the particular subject. They are usually graded on a weekly basis through quizzes or peer-graded assignments. The objective of this paper is to extract the concepts taught in the videos from the subtitles, which could later be used to enhance recommendations of the learners using their clickstream data. The teachers could also use this to see the demand for their courses. Evaluate two keyword extraction methods, which are BERT and LDA using different Coursera courses. The experimental results show that BERT outperforms LDA in terms of Coherence.