Decoding the Obfuscated: Advanced NER Techniques on Commercial Sex Advertisement Data

Detecting potential human trafficking activity within online Commercial Sex Advertisements (CSAs) presents unique challenges for Named Entity Recognition (NER) due to the complex and often obfuscated textual content. This paper thoroughly evaluates state-of-the-art language models and tokenization techniques, focusing on their efficacy in the NER task within the context of CSAs. Our results indicate that the Longformer model, equipped with byte-level BPE tokenization, outperforms other models regarding precision, recall, and the $F$1 score. The study also uncovers specific areas for improvement, offering avenues for future research. Our findings have significant implications for the automated analysis and monitoring of CSAs for suspected human trafficking activity, which contributes to developing more robust and transparent online ecosystems.