Mama/Papa, Is this Text for Me?

Proceedings of the 28th International Conference on Computational Linguistics, pages 6296–6301 Barcelona, Spain (Online), December 8-13, 2020 6296 Mama/Papa, Is this Text for Me? Md-Rashedur Rahman Gw´enol´e Lecorv´e Univ Rennes, CNRS, IRISAJonathan Chevelu Lannion, France first.last@irisa.fr Nicolas B´echet Universit´e Bretagne Sud Vannes, France first.last@irisa.fr Aline ´Etienne Delphine Battistelli Universit´e Paris-Nanterre, CNRS, MoDyCo Paris, France first.last@parisnanterre.fr Abstract Children have less linguistic skills than adults, which makes it more difficult for them to under- stand some texts, for instance when browsing the Internet. In this context, we present a novel method which predicts the minimal age from which a text can be understood. This method analy- ses each sentence of a text using a recurrent neural network, and then aggregates this information to provide the text-level prediction. Different approaches are proposed and compared to base- line models, at sentence and text levels. Experiments are carried out on a corpus of 1, 500 texts and 160K sentences. Our best model, based on LSTMs, outperforms state-of-the-art results and achieves mean absolute errors of 1.86 and 2.28, at sentence and text levels, respectively. 1 Introduction In recent years, safe Internet for children has gained interest in many research domains (Tomczyk and Kopeck`y, 2016; Byrne and Burton, 2017; Livingstone, 2019). However, most studies focus on abusive texts containing hate, violence, pornography, etc. (Liu and Forss, 2015; Suvorov et al., 2013). On the contrary, the adequacy of textual contents with the reading and understanding capabilities of children remains yet mainly unresolved in computational linguistics. Hence, this paper propose a new method to predicting this adequacy. Among the related works, (Schwarm and Ostendorf, 2005) explored the possibility of predicting from which US school grade newspaper articles could be read. This task was modelled as a classification problem among 4 classes using support vector machine fed with word-based n-gram probabilities, as well as lexical and syntactic features. (Islam and Rahman, 2014) has proposed a readability classifica- tion method for Bangla news articles for children. This method predicts if a text is either very easy, easy, medium or difficult. More recently, (Blandin et al., 2020) proposed different feed-forward (FF) neural models for age recommendation on texts targeting either children (from 0 to 14) or adults. The authors consider this as a regression task and explore various linguistic features and word embedding features, from which word embedding are shown as the most contributory. Overall, one can notice that these papers either rely on hand-crafted features or on simple models which consider texts as a global object rather than word sequences. This motivates us to further explore with word embeddings only and introduce recurrent neural networks (RNNs). More broadly, in the field of text readability, inherited from historical approaches like (Kincaid and Chissom, 1975), audiences other than children have been studied, e.g., second language learners (Xia et al., 2016), adults readers (Crossley et al., 2017) or patients interacting with doctors (Balyan et al., 2019). In parallel, text understanding by children is a well known question in psycho-linguistics and cognitive sciences. In particular, …

Content truncated to 500 words