... classificationtraining data. 2,000 articles containing smiles and2,000 articles containing frowns were held-out asoptimising test data. We took increasing amountsof articles from the remaining dataset ... 22,000 in increments of 1,000, an equal numberbeing taken from the positive and negative sets) asoptimising training data. For each set of trainingdata we extracted a context of an increasing ... Dependencies in Sentiment Classification2.1 Experimental Setup In this section, we describe experiments we havecarried out to determine the in uence of domain,topic and time on machine learning based...