Cooking Up a Neural-based Model for Recipe Classification

Proceedings of the 12th Conference on Language Resources and Evaluation (LREC 2020), pages 5000–5009 Marseille, 11–16 May 2020 c⃝ European Language Resources Association (ELRA), licensed under CC-BY-NC 5000 Cooking Up a Neural-based Model for Recipe Classification Elham Mohammadi1,2,3, Nada Naji1, Louis Marceau1, Marc Queudot1,3 Eric Charton1, Leila Kosseim2, Marie-Jean Meurs3 1Banque Nationale du Canada, 2Concordia University, 3Université du Québec à Montréal UQAM, Montreal, QC, Canada {nada.naji, louis.marceau, marc.queudot, eric.charton}@bnc.ca {elham.mohammadi, leila.kosseim}@concordia.ca meurs.marie-jean@uqam.ca Abstract In this paper, we propose a neural-based model to address the first task of the DEFT 2013 shared task, with the main challenge of a highly imbalanced dataset, using state-of-the-art embedding approaches and deep architectures. We report on our experiments on the use of linguistic features, extracted by Charton et al. (2014), in different neural models utilizing pretrained embeddings. Our results show that all of the models that use linguistic features outperform their counterpart models that only use pretrained embeddings. The best performing model uses pretrained CamemBERT embeddings as input and a Convolutional Neural Network (CNN) as the hidden layer, and uses additional linguistic features. Adding the linguistic features to this model improves its performance by 4.5% and 11.4% in terms of micro and macro F1 scores, respectively, leading to state-of-the-art results and an improved classification of the rare classes. Keywords: Embeddings, Neural Classification, Linguistic Features, Cooking Recipe Classification 1. Introduction In this paper, we present the neural models we developed to revisit the DEFT (Defi Fouille de Texte) 2013 (Grouin et al., 2013) shared task, using state-of-the-art embedding approaches and deep architectures. The DEFT 2013 shared task addressed the general task of mining cooking recipes in French through several classification and information ex- traction tasks. Task 1, which is the focus of this paper, ad- dressed the classification of recipes into four levels of diffi- culty (Very Easy, Easy, Fairly Difficult, and Difficult). The dataset is based on a collaborative website where individual users can post their own recipes. Due to this collaborative nature, the dataset raises several challenges, including the subjectivity of the labels and the use of varied vocabulary and grammar. Moreover, the dataset is highly imbalanced with approximately 90% of the samples belonging to the Very Easy and Easy classes, and only around 8% and less than 1% of the samples with a Fairly Difficult and Difficult label, respectively. These challenges sparked our interest and curiosity as to how the performances of traditional and state-of-the-art approaches would compare in a classifica- tion task. Additionally, even though the dataset revolves around food preparation, the challenge of an imbalanced dataset can be encountered in a variety of applications and real-world use cases. To address the challenge of a high class imbalance, we ex- perimented with pretrained embeddings as input features, and various deep architectures. We also experimented with adding the linguistic features extracted by Charton et al. (2014) to the deep models and measured per-class results to see if our approach could improve the performance on rare classes. This paper is structured as follows: The …

Content truncated to 500 words