| dc.description | The excessive use of filters on photos, along with the hundreds of profiles on social networks that abuse retouching to reduce centimeters of their bodies, and the existing social pressure on body image, has caused an increase in Eating Disorders (ED). Thus, anorexia, bulimia nervosa and binge eating disorder are the main eating disorders that put physical and mental health at risk, especially among the very young people. Fortunately, there are technologies that allow the early detection of certain problems in different areas, in particular those related to safety and health, such as the mentioned above. Our main objective in this work is to analyze how different representations of texts behave for the early detection of people suffering from anorexia. Although we focus on ED, we believe that these results could be extended to other risks such as depression, gambling, etc. We employ k-TVT, an efficient and effective method used previously in the detection of signs of depression, as well as other more elaborated approaches such as Word2Vec, GloVe, and BERT. To compare the performance of these methods, we worked on a data collection provided by the eRisk 2018 laboratory to detect signs of anorexia disorder. Regarding the results, the performance of the different approaches was quite similar, with k-TVT and BERT being slightly better. We also conclude that k-TVT continues to be efficient with its flexibility, low dimension and easy computing being the more attractive characteristics. | |
