Text classification to streamline online wildlife trade analyses

Area of Science:

• Ecology
• Conservation Biology
• Computer Science

Background:

• Automated monitoring of online wildlife trade is crucial for conservation and biosecurity.
• Identifying relevant advertisements on e-commerce sites is challenging due to unstructured text and large data volumes.
• Machine learning and natural language processing offer potential solutions for automated data extraction.

Purpose of the Study:

• To evaluate the effectiveness of text classifiers in extracting relevant wildlife trade advertisements from online platforms.
• To determine the minimum data sample size required for accurate performance of text classification models in this context.

Main Methods:

• Collected 16.5k pet bird advertisements from an Australian classifieds website.
• Applied a suite of text classifiers to identify relevant listings.
• Conducted a sensitivity analysis by reducing sample sizes to assess model performance.

Main Results:

• Text classifiers achieved high accuracy in identifying relevant wildlife trade listings (ROC AUC ≥ 0.98, F1 score ≥ 0.77).
• A minimum sample size of 33% (approx. 5.5k listings) was found to be necessary for adequate model performance.
• Text classification significantly reduces time spent on data cleaning for wildlife trade monitoring.

Conclusions:

• Text classification is a viable tool for processing online wildlife trade data, aiding conservation efforts.
• Model performance is context-dependent, varying with advertisements and websites.
• Integrating image classification with text classification could further enhance predictive accuracy and data processing efficiency.