Introduction
The Illegal Wildlife Trade (IWT) of species (flora and fauna) is having a direct impact on endangered species worldwide, putting at risk biodiversity and environmental sustainability.
In the plant domain poachers are routinely destroying vast habitats for CITES protected species, such as rare cacti which take 100+ years to grow in the wild, which once gone are irreplaceable. Poachers will traffic illegally obtained wildlife to low regulation third party countries (e.g. China) where wildlife is sold via online auction sites such as Alibaba, eBay and Amazon using euphemisms and trade jargon to avoid online platform detection and takedown of the adverts. Law enforcement (e.g. UK Border Force) need AI to efficiently monitor IWT for target species to allow better intelligence led interventions.
This PhD project will explore how to use Multimodal Large Language Models to perform Information Extraction on auction HTML web pages obtained from web crawling software (e.g. MEMEX crawler).
LLMs have recently been shown [1] effective at automatically parsing HTML for information extraction. This PhD will explore human-in-the-loop LLM training ideas (e.g. deep active learning, rationale-based training, multi-agent debating, LLM preference alignment) allowing efficient LLM model fine-tuning through a set of feedback sessions that conservation experts could participate in. The downstream application goal is to develop LLM algorithms that can be used with a crawler to build a IWT intelligence picture about what species are being traded, by whom and where.
An inter-disciplinary project, a conservation expert from Kew Gardens will join the team as an external supervisor to provide IWT expertise. If successful, results can be presented to contacts within the UK's National Wildlife Crime Unit and UK Border Force for potential tool adoption by law enforcement officers.
References:
References
[1] https://serpapi.com/blog/traditional-parsing-vs-mistral-7b/