With the growth that social networks have experienced in recent years, it is entirely impossible to moderate content manually. Thanks to the different existing techniques in natural language processing, it is possible to generate predictive models that automatically classify texts into different categories. However, a weakness has been detected concerning the language used to train such models. This work aimed to develop a predictive model based on BERT, capable of detecting racist and xenophobic messages in tweets written in Spanish. A comparison was made with different Deep Learning models. A total of five predictive models were developed, two based on BERT and three using other deep learning techniques, CNN, LSTM and a model combining CNN + LSTM techniques. After exhaustively analyzing the results obtained by the different models, it was found that the one that got the best metrics was BETO, a BERT-based model trained only with texts written in Spanish. The results of our study show that the BETO model achieves a precision of 85.22% compared to the 82.00% precision of the mBERT model. The rest of the models obtained between 79.34% and 80.48% precision. On this basis, it has been possible to justify the vital importance of developing native transfer learning models for solving Natural Language Processing (NLP) problems in Spanish. Our main contribution is the achievement of promising results in the field of racism and hate speech in Spanish by applying different deep learning techniques.
BACKGROUND: Eating disorders affect an increasing number of people. Social networks provide information that can help. OBJECTIVE: We aimed to find machine learning models capable of efficiently categorizing tweets about eating disorders domain. METHODS: We collected tweets related to eating disorders, for 3 consecutive months. After preprocessing, a subset of 2000 tweets was labeled: (1) messages written by people suffering from eating disorders or not, (2) messages promoting suffering from eating disorders or not, (3) informative messages or not, and (4) scientific or nonscientific messages. Traditional machine learning and deep learning models were used to classify tweets. We evaluated accuracy, F1 score, and computational time for each model. RESULTS: A total of 1,058,957 tweets related to eating disorders were collected. were obtained in the 4 categorizations, with The bidirectional encoder representations from transformer-based models had the best score among the machine learning and deep learning techniques applied to the 4 categorization tasks (F1 scores 71.1%-86.4%). CONCLUSIONS: Bidirectional encoder representations from transformer-based models have better performance, although their computational cost is significantly higher than those of traditional techniques, in classifying eating disorder-related tweets.
