Detection of Anorexic Girls-In Blog Posts Written in Hebrew Using a Combined Heuristic AI and NLP Method

In this study, we aim to detect in social media texts written in Hebrew girls who are suspected of being anorexic. We constructed a dataset containing 100 blog posts written by females who are probably anorexic, and 100 blog posts written by females who are likely to be non-anorexic. The construction of this dataset was supervised and approved by an international expert on anorexia. We tested several text classification (TC) methods, using various feature sets (content-based and style-based), five machine learning (ML) methods, three RNN models, four BERT models, three basic preprocessing methods, three feature filtering methods, and parameter tuning. Several insights were found as follows. A set of 50-word n-grams (mostly word unigrams) given by an expert was found as a good basic detector. A heuristic process based on the random forest ML method has overcome a combinatorial explosion and led to significant improvement over a baseline result at a level of
$\text{P}\,{=}$
.01. Application of an iterative process that tests combinations of “k out of
$\text{n}'$
” where
$\text{n}'\,{ < }$
n (n is the number of feature sets) lead to a result of 90.63%, using a combination of 300 features from ten feature sets.