Pencil

Classifying human rights violations with deep multi-label co-training.

Evolution. Blum’s Co-training – Running two classifiers over a given view to learn and classify content given a small labelled set Multi-labelled co-training – Co-training was extended to support multi-labels Content based Co-training – Co-training was extended to use matrix factorization Co-training with recommendations – To solve cold start problem Deep multi-label co-tranining.

Method. Uses two neural networks to train on different views of generated samples to calculate similarity in the probability distribution of predicted outcomes. Cumulative loss Multi-label sigmoid cross-entropy – for the multi label classification Kullback-Leibler divergence - to find out how the two view differ from each other Noise sigmoid cross-entropy – Regularizing the added noise Adds noise to keep the classifier from being affected by it during prediction. words with given probability; replacing words by a filler token with given probability; swapping words up to a certain range..

Results. 10% of labelled set. o-t Ikrvised multi-la Tree SVM Naive Baves o-training multi-lab Decision Tree SVM Naive Baves n ng nuj ecas.on ree KNN SVM Naive Bayes x acto za n Co-training multi-label with similarity recommendations Decision Tree KNN SVM Naive Bayes mu t NN co-tr accuracy 0.71015873015873 0336796536796537 0.71448773"87735 0.71015873015873 0.4321548821548822 4.9090'*j909fM91E+016 0.4993265993265993 4.93434343434343E+016 0.7057239057239€Y, 0.695622895622896 0.730832130832131 0.7663343 0.6299807 0.72343244 6 error rate 0.27984126984127 0.463203463203463 0265512265512265 0.28984126984127 0.5678451 178451178 0.5065656565656567 O.304377104377104 0.269167869167869 0.2336657 0.3700193 0.2432(MY787 0.27656756 TABLE I: Classifier comparison on of the labeled samples.

Results. With different thresholds of the labelled set.

Future. The present study used the domain of human rights violations. However can be applicable where - The aim is to identify and ignore forms of misinformation has an initial small labelled Has minimal set of experts.