Accuracy can be improved using one or more of these strategies.
Text Cleaning. Removing stopwords, punctuations marks, suffix variations etc. can reduce the noise present in text data.
Hstacking arrays of text / NLP features with text feature vectors in sequence horizontally (column wise). After extracting features initially, try combining different feature vectors to improve the accuracy of the classifier.
Tuning the Classifier. Adjusting the parameters of the classifier is an important step towards a best fit model.
Testing Different Ensemble Models. Stacking different models and blending their outputs can help to further improve the results.
Weighting words: we can significantly improve our algorithm by accounting for the commonality of each word. The word “is” should carry a lower weigh than the word “sandwich” in most cases, because it is more common.
How to evaluate the different components of the overall text classification problem.
One of the most common methods for quantifying the discrimination level of a feature is the use of a measure known as the gini-index.
Cohen’s kappa is a measure of interrater reliability used in social sciences. This excerpt from an R blog explains it.
Using kappa score for comparisons of groups and methods https://www.personality-project.org/r/html/kappa.html https://www.r-bloggers.com/k-is-for-cohens-kappa/
Last April, during the A to Z of Statistics, I blogged about Cohen’s kappa, a measure of interrater reliability. Cohen’s kappa is a way to assess whether two raters or judges are rating something the same way. And thanks to an R package called irr, it’s very easy to compute. But first, let’s talk about why you would use Cohen’s kappa and why it’s superior to a more simple measure of interrater reliability, interrater agreement.
Copyright © 2019 A. Daniel Johnson. All rights reserved.