TLDR: This research explores neurosymbolic methods, combining large language models (LLMs) with formal reasoning, to identify forced labor in complex supply chains. It details manual and automated feature extraction from news articles using a question tree framework and proposes Boolean formula enumeration to find patterns indicative of illicit activity, aiming to improve detection and inform policy.
Global supply chains are incredibly intricate, making them challenging to monitor, especially when illicit activities like forced labor, human trafficking, or counterfeit goods are involved. Traditional machine learning (ML) methods often fall short in these scenarios because they require vast amounts of training data, which is typically sparse, corrupted, or intentionally hidden in illicit supply chains. A new research paper introduces a novel approach using neurosymbolic methods to automatically detect patterns linked to illegal activities, even with limited and unreliable data.
The paper, titled “Neurosymbolic Feature Extraction for Identifying Forced Labor in Supply Chains,” by Zili Wang, Frank Montabon, and Kristin Yvonne Rozier from Iowa State University, explores how to identify instances of illicit activity, specifically forced labor, in supply chains. Their work compares the effectiveness of both manual and automated feature extraction from news articles that describe illicit activities uncovered by authorities. A key innovation is their proposed “question tree” approach, which queries a large language model (LLM) to identify and quantify the relevance of articles, allowing for a systematic evaluation of how humans and machines classify news related to forced labor.
Understanding the Approach
The core of this research lies in combining the pattern-recognition capabilities of large language models (the “neuro” part) with the precision and interpretability of formal logic (the “symbolic” part). The goal is to extract meaningful indicators, or features, from publicly available information like news articles, which can then be analyzed to detect forced labor.
How Data is Extracted
The researchers employed two main methods for extracting data:
Manual Feature Extraction: To build a foundational dataset, human experts queried online news databases like ProQuest and LexisNexis using terms such as “forced labor” and “supply chain.” From 2016 to 2024, over 340 articles were gathered. These articles were then manually classified as relevant or irrelevant, and 25 specific features indicative of forced labor were extracted from the relevant ones. This process resulted in 125 documented incidents across various industries, including textiles, seafood, agriculture, and precious metals. For example, an incident involving Chinese tuna fishing vessels using North Korean forced laborers was identified, with features like “tuna” as the product, “seafood” as the supply chain, and “China” as the country of incident.
Automated Feature Extraction: To scale this process, the researchers leveraged the GPT-4.0 large language model. The LLM was prompted to search for articles related to forced labor in supply chains by querying for relevant keywords via the ProQuest API. A crucial component of this automated method is the “question tree framework.” This framework is a structured set of questions designed to evaluate an article’s relevance to forced labor. Starting with a root question like “Does the article mention forced labor?”, the LLM proceeds through a series of interconnected questions. A positive answer to one question can lead to further, more specific questions. Each positive answer contributes to a relevance score for the article, allowing for automated classification.
Analyzing the Relationships Between Features
Once features are extracted, whether manually or automatically, the next step is to understand how they relate to each other in the context of forced labor. The paper proposes using a SAT-based Boolean formula enumeration technique. This method encodes the extracted features as Boolean variables (true/false) and then systematically identifies combinations of these features that are highly indicative of forced labor. For instance, their previous work identified a formula: “cross_border ∧ (high_risk_source ∨ high_risk_product).” This suggests that if a product crosses a national border AND originates from a high-risk country OR is a high-risk product itself, it is a strong indicator of potential forced labor.
To fully utilize this technique, the researchers note the need for data points representing *non-instances* of forced labor, which would allow for a comparison to determine how meaningful a formula is. One potential solution is to use the same LLM to classify and extract features from articles initially deemed irrelevant, thus creating a dataset of non-instances.
Also Read:
- SymbolicThought: Enhancing Narrative Understanding with AI and Logic
- KEA Explain: A New Approach to Detecting and Explaining AI Hallucinations
Looking Ahead
The researchers envision several future directions for this work. They aim to improve data collection by adding more articles and enhancing quality, possibly by having multiple human experts or an ensemble of LLMs classify the same articles. Combining manual and automated methods is also a key focus, allowing human domain knowledge to refine the automated processes. Furthermore, they plan to expand the use of formal methods to include temporal or epistemic features, using logics like Mission-time Linear Temporal Logic (MLTL) to detect patterns over time, such as unusual delays between recruitment and the start of work.
Ultimately, this research aims to foster wider adoption of neurosymbolic methods in supply chain analysis and other domains plagued by illicit activities. The insights gained are intended to inform legislation, help companies reduce compliance costs, guide law enforcement efforts, and significantly reduce the global prevalence of forced labor in supply chains. You can read the full research paper here: Neurosymbolic Feature Extraction for Identifying Forced Labor in Supply Chains.
