Context-specific Language Modeling for Human Trafficking Detection from Online Advertisements

Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 1180–1184 Florence, Italy, July 28 - August 2, 2019. c⃝2019 Association for Computational Linguistics 1180 Context-specific language modeling for human trafficking detection from online advertisements Saeideh Shahrokh Esfahani Accenture Technology Labs San Francisco, CA saeideh.shahrokh@accenture.com Michael J. Cafarella Department of Computer Science University of Michigan michjc@umich.edu Maziyar Baran Pouyan Accenture Technology Labs San Francisco, CA maziyar.baran.pouyan@accenture Gregory J. DeAngelo Department of Economics Claremont Graduate University gregory.deangelo@cgu.edu Elena Eneva Accenture Technology Labs San Francisco, CA elena.eneva@accenture.com Andrew E. Fano Accenture Technology Labs San Francisco, CA andrew.e.fano@accenture.com Abstract Human trafficking is a worldwide crisis. Traf- fickers exploit their victims by anonymously offering sexual services through online adver- tisements. These ads often contain clues that law enforcement can use to separate out po- tential trafficking cases from volunteer sex ad- vertisements. The problem is that the sheer volume of ads is too overwhelming for man- ual processing. Ideally, a centralized semi- automated tool can be used to assist law en- forcement agencies with this task. Here, we present an approach using natural language processing to identify trafficking ads on these websites. We propose a classifier by integrat- ing multiple text feature sets, including the publicly available pre-trained textual language model Bi-directional Encoder Representation from transformers (BERT). In this paper, we demonstrate that a classifier using this com- posite feature set has significantly better per- formance compared to any single feature set alone. 1 Introduction In 2013, the Global Slavery Index reported that 30 million individuals were living in involuntary servitude. Another estimation found that 600,000 women are trafficked in the sex industry per year with the United States being the second most popular destination for these individuals (Kara, 2009); (Schauer and Wheaton, 2006). In the last decade, it has become more difficult for law en- forcement (LE) to trace traffickers as they have begun to take increasing advantage of online ad- vertisement platforms for sexual services to solicit clients and become less visible. LE is capable of tracking the posted ads and mining such data to detect trafficking victims. However, the large vol- ume of online unstructured data, the high degree of similarity of ads (Figure 1), and the lack of an automated approach in detecting suspicious activ- ities through advertisements present obstacles for LE to independently develop methods for survey- ing these criminal activities. Sex trafficking ad- vertisements are unique texts. They have incor- rect grammatical structures and misspellings, and are enriched with unconventional words, abbre- viations, and emojis. Oftentimes the author uses emojis and emoticons to convey messages to a po- tential customer. In particular these types of ad- vertisements may also contain equivocal words, e.g., roses as a substitute for dollars. Addition- ally, dominant keywords from these online ads continuously evolve as traffickers and consent- ing sex workers alike seek to evade prosecution. While previous researchers have tried to develop automated systems to detect trafficking advertise- ments, this has proved an enormous challenge for natural language processing …

Content truncated to 500 words