Zero-Shot Turkish Text Classification

The method frequently used for text classification is supervised modeling with a large training set with labels. In some cases, we may not have labeled data. Modeling in the absence of labeled data for target classes is called zero-shot modeling. For zero-shot text classification natural language inference is utilized which is another branch of natural language processing. In order to measure the effectiveness of zero-shot classification, a reference is generated with BERT method which contrarily uses labeled data. Two Turkish data sets are employed. TTC-3600 dataset is composed of 6 news categories. In this study we have generated a new Twitter sentiment dataset TTD, which is composed of weakly labeled Twitter messages in 2 classes as positive and negative depending on contained emoticons. The results show that without using any labeled data we can obtain classification accuracies of %65.5 and %91 for TTC-3600 and TTD respectively.