IUU+DB: Tracking Illegal, Unreported, and Unregulated Fishing, Seafood Fraud, and Labor Abuse through LLM-driven Information Extraction

arXiv:2606.18181 · cs.IR, cs.AI, cs.CY · Submitted 2026-06-16 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Next we'll be talking about the paper "IUU+DB: Tracking Illegal, Unreported, and Unregulated Fishing, Seafood Fraud, and Labor Abuse through LLM-driven Information Extraction".

Jane: The paper was written by the authors from Virginia Tech and Simeone Consulting, LLC and Natural Resources Defense Council and University of Washington.

Tom: Stay tuned as we take you through the paper and discuss its implications.

Title: Tom: We are starting with "IUU+DB: Tracking Illegal, Unreported, and Unregulated Fishing, Seafood Fraud, and Labor Abuse through LLM-driven Information Extraction".

Jane: That title is quite a mouthful, Tom.

Tom: It really is, but it describes exactly what Henry Bodwell and his team are doing.

Jane: They are expanding the definition of illegal fishing, aren't they?

Tom: Yes, they are using this "IUU+" label to include things like seafood fraud.

Jane: And even labor abuse within those same supply chains.

Tom: This is a much broader way of looking at maritime crime.

Jane: It covers so many different types of misconduct.

Tom: Exactly, and that's why the "plus" is so important.

Lu: I think it's beautiful how this connects environmental damage with human rights.

Meng: I do wonder if the system can handle all those different categories without getting confused.

Jane: That is the big challenge they are tackling, Meng.

Lu: This is like building a digital net that can catch even the most subtle connections.

Meng: If they can get the data right, it's a huge win for investigators.

Lalam: This research could change our entire cultural perception of seafood ethics.

Tom: It moves the focus from just boats to entire ecosystems.

Jane: I agree, it's a much more holistic view.

Tom: We should look at how they actually built the engine next.

Summary: Jane: We've seen the vision, so let's talk about the mechanics of "IUU+DB: Tracking Illegal, Unreported, and Unregulated Fishing, Seafood Fraud, and Labor Abuse through LLM-driven Information Extraction".

Tom: They have to turn a mountain of messy text into structured data.

Jane: They pull from news articles and government reports.

Meng: How do they deal with documents that are just images, like old PDFs?

Tom: They use a tool called Pytesseract to perform OCR.

Jane: This turns those images into plain text that the AI can read.

Meng: That sounds like a lot of heavy lifting for the preprocessing stage.

Tom: It is, and they even use Qwen embeddings to store it in a vector database.

Jane: This allows the system to understand the meaning behind the words.

Lu: I think using that kind of high-dimensional space to organize crime is brilliant.

Meng: But how do they extract over a hundred specific details without errors?

Tom: They use a framework called DSPy to guide the process.

Jane: They group those hundred fields into fourteen different categories.

Lu: This keeps the model focused on one specific area at a time.

Meng: So it's solving many small puzzles instead of one giant one.

Lalam: They are essentially teaching a machine to be a specialized investigator.

Tom: This is a very sophisticated way to turn noise into a map.

Jane: I wonder if that map actually works in the real world.

Tom: That is exactly what we will find out in the next segment.

Improvements: Tom: We are looking at the results for "IUU+DB: Tracking Illegal, Unreported, and Unregulated Fishing, Seafood Fraud, and Labor Abuse through LLM-driven Information Extraction".

Jane: The scale is truly impressive, covering over eight thousand incidents.

Tom: They tracked data from more than one hundred and forty countries.

Meng: That is a massive amount of data for an automated system to process.

Jane: And they saw a fifteen to twenty percent improvement over standard models.

Tom: They compared their performance against GPT-4o Mini.

Meng: That jump in reliability is huge for anyone using this for enforcement.

Jane: It definitely makes the database more useful for real decisions.

Tom: They did note that some details like species or vessels can still be tricky.

Lu: The Maximum Entropy approach they used was my favorite part.

Meng: Is that the part that finds the statistical outliers?

Lu: Yes, it flags events that are behaving strangely for a specific region.

Jane: Like that unexpected spike in fishing conflicts in Sri Lanka.

Tom: Or the case in the US involving undersized fish.

Lalam: This gives the system a kind of digital intuition for spotting trouble.

Meng: That's a proactive way to monitor global activity.

Lu: This turns a database into a dynamic warning system.

Jane: This helps us find the signal in the noise.

Tom: And that's where the real value lies.

Conclusion: Tom: We are wrapping up our discussion on "IUU+DB: Tracking Illegal, Unreported, and Unregulated Fishing, Seafood Fraud, and Labor Abuse through LLM-driven Information Extraction".

Jane: This research shows how AI can become a real tool for global accountability.

Lu: I see this as the start of a permanent watchdog for our planet.

Lu: We could eventually see these systems monitoring every corner of the ocean.

Meng: I can see port inspectors and regulators using this to flag high-risk shipments.

Meng: It would save them so much time during inspections.

Lalam: It will help bring much-needed transparency to the food supply.

Lalam: This kind of visibility changes how we value what we eat.

Tom: This research focuses on making sure these complex systems are held accountable.

Jane: I agree, Tom, we need that oversight.

Tom: We also need to ensure we are protecting the people in the supply chain.

Jane: That is so important.

Jane: It has been a fascinating look at how technology meets environmental justice.

Lu: I am really looking forward to the next version of this.

Meng: I will be watching for the practical applications.

Lalam: This is a major step toward a more ethical world.

Tom: Thanks to the whole team for joining us.

Jane: We'll be back very soon.

Tom: Next time, we're diving into the world of quantum computing.

Virginia Tech · Simeone Consulting, LLC · Natural Resources Defense Council · University of Washington

cs.IR, cs.AI, cs.CY

Submitted: 2026-06-16

Updated: 2026-09-12

Project page: https://dspy.ai

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 81/100

The gist: This paper introduces IUU+DB, an LLM-driven information extraction framework designed to build a global incident database for illegal, unreported, and unregulated (IUU) fishing, seafood fraud, and

Key concepts

IUU+ Label: This expands the traditional definition of illegal, unreported, and unregulated fishing to include seafood fraud and labor abuse within supply chains. It provides a holistic view of maritime misconduct by connecting environmental damage with human rights issues across entire ecosystems rather than just focusing on individual boats.
Information Extraction Framework: The system uses Pytesseract for OCR to convert images into text and the DSPy framework to extract over one hundred specific details. These details are organized into fourteen categories, allowing the AI to act like a specialized investigator that turns messy text into structured, usable data.
Maximum Entropy Approach: This method identifies statistical outliers within the database to flag unusual activity. By detecting events that behave strangely for a specific region—such as unexpected spikes in fishing conflicts or reports of undersized fish—it serves as a dynamic warning system for global monitoring and enforcement.

Terminology

Adam — 7× in this paper · explained in 2 episodes
Actor — 6× in this paper · explained in 3 episodes
Aggregation — 4× in this paper · explained in 1 episode
Core — 3× in this paper · explained in 2 episodes
Seafood Import Monitoring Program — 3× in this paper
Accuracy — 2× in this paper · explained in 1 episode
Cosine Similarity — 2× in this paper · explained in 3 episodes
Hallucination — 2× in this paper · explained in 4 episodes
Ablation — 1× in this paper · explained in 1 episode
Accountability — 1× in this paper · explained in 2 episodes
Consistency — 1× in this paper · explained in 3 episodes
Embeddings — 1× in this paper · explained in 2 episodes
F1 Score — 1× in this paper · explained in 3 episodes
Fine-Tuning — 1× in this paper · explained in 5 episodes
Grounding — 1× in this paper · explained in 3 episodes
Large Language Model (LLM) — 1× in this paper · explained in 3 episodes
Ontology — 1× in this paper · explained in 2 episodes
Optical Character Recognition — 1× in this paper

Summary

This paper introduces IUU+DB, an LLM-driven information extraction framework designed to build a global incident database for illegal, unreported, and unregulated (IUU) fishing, seafood fraud, and labor abuse. It addresses the difficulty of obtaining a quantitative understanding of these crimes by aggregating fragmented evidence from heterogeneous sources—such as news media, government announcements, academic publications, and NGO reports—to support targeted enforcement efforts to government agencies.

The IUU+ Taxonomy

The authors propose the term IUU+ to capture a broader suite of fisheries sector environmental and associated supply chain trade-related crimes and behaviors than traditional definitions. This inclusive taxonomy is organized into seven primary themes:

Illegal Fishing (e.g., exceeding catch quotas or using prohibited gear)
Unreported Fishing (e.g., misreporting target catch weight or species)
Unregulated Fishing (e.g., fishing without a flag state)
Seafood Fraud or Mislabeling (e.g., species substitution)
Forced Labor or Labor Abuse (e.g., wage violations or physical violence)
Circumventing Prohibitions or Sanctions (e.g., tariff evasion)
Aquaculture-Specific Activities (e.g., illegal sourcing of broodstock)

System Architecture and Methodology

IUU+DB functions as an end-to-end pipeline for turning unstructured documents into structured incident records. The process begins with a Document Pre-Processing Module that utilizes Pytesseract for OCR and Playwright for URL text extraction. A Source Classification Module then categorizes documents as describing an incident, being related to IUU+ without describing a specific incident, or being unrelated. Functionally, the system sets its minimum information threshold based on the common data elements used in crime script analysis: who (actors), what (methods), when (time), and where (location).

The KDE Extraction Module is the workhorse of the IUU+DB system, utilizing LLMs and DSPy to extract information across 14 groups of Key Data Elements (KDEs) covering over 100 possible fields. To manage this complex task, the pipeline first identifies which KDE groups are present in a document before attempting extraction, which effectively reduces the schema information the LLM needs to hold in context. Finally, a Deduplication and Trend Identification module uses cosine similarity and matching identifiers—such as vessel name, or id—to merge duplicate incidents and ensure event reconciliation.

Experimental Results and Evaluation

The framework ingested data from 2,472 sources across 143 countries, yielding 8,435 incidents. Through Hotspot Analysis, the system identified significant trends like illegal red snapper harvesting in the Gulf of Mexico. Using a maximum entropy approach, the researchers also discovered surprising country-behavior events, such as cases of falsifying tracking data in Saya de Malha.

In performance evaluations, IUU+DB demonstrated significant utility despite some classification challenges. While the incident classifier had a low precision because it was tuned to avoid missing incidents, the system generally performed well. Comparative analysis showed that IUU+DB offered a 15-20% improvement over the plain GPT-4o calls and 10% improvement over the newer GPT-5.4 model regarding IUU Type and Sub-Behavior classification tasks.

Improvements for AI systems

1. Implementation of Hierarchical, Group-Based Information Extraction (IE)

The Improvement: Move away from flat large-scale schema extraction toward a multi-stage pipeline that first identifies relevant semantic clusters (KDE groups) before attempting field-level extraction.
What the improved system can do: It can process extremely complex, high-dimensional schemas (e.g., 100+ data fields) without overwhelming the LLM's context window or increasing hallucination rates. By narrowing the focus to specific logical groups (e.g., Vessel Information or Labor Standards) before extracting individual fields, the system achieves higher precision and significantly lower computational latency in domain-specific environments.

2. Programmatic Prompt Optimization via Semantic Reward Functions

The Improvement: Replace manual prompt engineering with an automated optimization framework (such as DSPy’s MIPROv2) that utilizes a reward metric based on the cosine similarity between extracted values and validated ground truth, paired with negative reinforcement for hallucinations.
What the improved system can do: It can autonomously fine-tune its own instructions and field definitions. This allows the AI to self-correct and adapt to shifting linguistic nuances in heterogeneous source documents (e.g., moving from formal government reports to colloquial news articles) without human intervention.

3. Hybrid Consensus-Based Deduplication Engine

The Improvement: Integrate a dual-pass entity resolution system that combines semantic embedding similarity with hard-attribute matching (e.g., vessel IDs, timestamps, and location codes), supplemented by an automated adjudication query.
What the improved system can do: It can accurately merge duplicate incident reports from fragmented sources while resolving data conflicts. If two sources provide conflicting information (e.g., different arrest dates for the same vessel), the system will automatically trigger a targeted re-query of all attached source documents to establish a consensus, ensuring high data integrity for downstream trend analysis.

4. Multimodal Semantic Embedding Ingestion

The Improvement: Transition from text-only processing to a multimodal embedding architecture (using models like Qwen3-VL) that treats text and embedded visual data (tables, maps, and figures) as a unified semantic space.
What the improved system can do: It can extract structured data from non-textual elements within documents, such as capturing geographic coordinates from maps of illegal fishing zones or extracting specific values from complex tables in academic papers, which are currently invisible to standard text-based NLP pipelines.

Abstract

Illegal, unreported, and unregulated fishing (IUU) traditionally refers to fishing activities that violate applicable laws or occur in areas that lack applicable laws. We propose the term IUU+ to capture a broader suite of fisheries sector environmental and associated supply chain trade-related crimes and behaviors. Although IUU+ activity is widely recognized as a serious threat to marine ecosystems, markets, and livelihoods, a quantitative understanding of these incidents, e.g., their frequency, geography, species, actors, and patterns in the type of illicit activity, remains difficult to obtain. We propose IUU+DB, a large language model driven system for building a global incident database of IUU+ activity. The system ingests heterogeneous documents, classifies whether they describe relevant incidents, extracts key data elements such as actors, locations, species, vessels, violations, and enforcement outcomes, and supports deduplication and trend analysis. Case studies and validation results show that IUU+DB can help organize fragmented evidence, surface geographic and behavioral hotspots, support fisheries-domain specific research in academia and non-government organizations, assist source and species risk assessments for industry, and provide support for policy implementation and targeted enforcement efforts to government agencies.