Classification of Rare Recipes Requires Linguistic Features as Special Ingredients

Elham Mohammadi, Nada Naji, Louis Marceau, Marc Queudot, Eric Charton, Leila Kosseim, Marie-Jean Meurs

Abstract: In this paper, we propose a joint model, composed of neural and linguistic sub-models, to address classification tasks in which the distribution of labels over samples is imbalanced. Different experiments are performed on tasks 1 and 2 of the DEFT 2013 shared task [10]. In one set of experiments, the joint model is used for both classification tasks, whereas the second set of experiments involves using the neural sub-model, independently. This allows us to measure the impact of using linguistic features in the joint model. The results for both tasks show that adding the linguistic sub-model improves classification performance on the rare classes. This improvement is more significant in the case of task 1, where state-of-the-art results are achieved in terms of micro and macro-average F1 scores.