Reference no: EM133217004

Question 1.

Imagine you have been assigned the task of building a classifier for sentiment analysis. The classifier must label texts at a sentence level, giving each sentence a label of positive, negative or neutral. The data to be classified consists of multi-sentence documents of variable lengths, in English. You are provided with a labelled dataset of 1,000 documents, containing an approximately equal number of each of the three classes (positive, negative, neutral). You also have access to large amounts of unlabelled English text.

1. Describe a suitable neural network architecture for this task, explaining the role of each layer and its inputs and outputs, and justifying your choices of network structure, representation and activation functions.

2. Explain how you could use the labelled and unlabelled data to train your classifier, including choice of loss function. (You don't need to give mathematical formulae for loss or activation functions.)

Question 2.

Now imagine that the labelled dataset in part (a) above did not contain an even class distribution, but was unbalanced, with 700 documents labelled neutral, 50 positive and 250 negative. Explain what consequences this would have for your choice of loss function in training, and of evaluation metric when testing.

Question 3.

Neural network classifiers in settings like this can run the risk of over-fitting. Explain the concept of over-fitting and how you could go about detecting it, and discuss what steps you could take to mitigate the problem.