Copy-Pasted Comments Dataset

5.1. Dataset As we know Deep Neural Networks are data intensive and it performs better if trained with a dataset with large amount of data. As dataset in Bangla language is publicly not available, we collect about 850 Bangla comments from different sources and process them to make them suitable to train the network. Among the comments, 500 comments are positive and the rests are negative. But as the amount of comments is not sufficient to train our proposed network, we just copy-paste the comments repeatedly to increase the size of our dataset. We increase the number of comments in our dataset approximately to 120000. Where 60000 comments are positive and 60000 comments are negative. We divide the positive and negative comments into separate files. To evaluate the model at training period we use 10-fold cross-validation on the training data.

A screenshot of a dataset section from a technical paper explains that only 850 Bangla comments were collected, then repeatedly copy-pasted to create a dataset of approximately 120000 comments. The joke is the absurdly unsound machine-learning methodology: duplicated comments are presented as if they meaningfully increase the training data. It is used as a reaction to fake data science, questionable research, or effortless attempts to inflate results.

• @profunctor_io / 4940 2020-07-15