💬 NLP

Peacemaker at ATE-IT: Automatic term extraction from Italian text for waste management data using encoder model

This paper presents a low-cost, interpretable automatic term extraction method for Italian waste management data, which utilizes fine-tuning strategies to achieve balanced performance on the ATE-IT Task A despite limited annotated resources.

Original authors: Mahdi Bakhtiyarzadeh, Hadi Bayrami Asl Tekanlou, Jafar Razmara

Published 2026-06-02

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Mahdi Bakhtiyarzadeh, Hadi Bayrami Asl Tekanlou, Jafar Razmara

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to teach a computer to read a very specific, boring, and complex manual about trash collection and waste management in Italy. Your goal is to get the computer to highlight the most important "keywords" or "phrases" (like "waste collection service" or "public health protection") so that it can understand the document better. This task is called Automatic Term Extraction (ATE).

The authors of this paper, a team from the University of Tabriz called "Peacemaker," built a tool to do exactly this for the Italian language. Here is how they did it, explained simply:

1. The Challenge: Finding the "Needles" in the "Haystack"

The team was given a pile of Italian legal and administrative texts about waste management. They needed to find specific phrases that act as the "heartbeats" of these sentences.

The Problem: It's hard for computers to know where a phrase starts and ends, especially when words are long and complex. Also, they didn't have a massive library of pre-labeled examples to teach the computer, so they had to be efficient.
The Goal: Create a system that is cheap to run, easy to understand, and accurate enough to find these terms without needing a supercomputer.

2. The Solution: A "Smart Highlighter"

Instead of building a giant, complicated robot, the team built a lightweight, smart highlighter.

The Brain (The Model): They used a pre-trained AI brain called BERT (specifically a version trained on Italian). Think of this as a student who has already read millions of Italian books and understands the grammar and flow of the language perfectly.
The Training (Fine-Tuning): They didn't teach the student everything from scratch. Instead, they gave the student a specific set of practice tests (the waste management texts) and said, "Look, this is a term, and that is just a regular word." The student learned to recognize the pattern.
The Method (BIO Tagging): To make this work, they turned the task into a game of "tagging."
- B (Begin): The first word of a term gets a "Start Here" tag.
- I (Inside): The middle words of a term get a "Keep Going" tag.
- O (Outside): Words that aren't part of a term get an "Ignore" tag.
- Analogy: Imagine a line of people. The team taught the computer to put a red hat on the first person of a group, a blue hat on the people in the middle of the group, and leave everyone else bare-headed.

3. The Rules of the Game (The Competition)

The team entered a competition called ATE-IT 2026, where 9 different teams tried to solve this puzzle.

The Rules: They had to find the terms in a "Test" set of sentences that they hadn't seen before.
The Scorecard: They were judged on two things:
1. Precision: When you said "This is a term," were you right? (Did you avoid false alarms?)
2. Recall: Did you find all the terms that were actually there? (Did you miss any?)
3. Two Levels: They checked this on a "Type" level (did you find the types of words?) and a "Micro" level (did you find the exact instances of words in every sentence?).

4. The Results: The "Underdog" Story

The Peacemaker team didn't have the biggest budget or the most powerful computers. They used a modest setup.

The Outcome: They finished 7th out of 9 teams.
The Twist: While they didn't win first place, they were very proud of their consistency.
- Analogy: Imagine a race where some runners sprinted fast but stumbled often, while others jogged steadily. The Peacemaker team was like the steady jogger. They didn't have the highest speed, but they didn't make many mistakes, and they found both common and rare terms with equal care.
Why it matters: They proved that you don't need a massive, expensive supercomputer to get decent results. A simple, well-tuned model can do a solid job, making this technology accessible to smaller organizations.

5. What They Didn't Do

The paper is very honest about its limits:

They did not use AI to write the data or the answers; humans did the hard work of labeling.
They did not claim their system is perfect or ready to replace human experts immediately.
They did not promise that this will fix the world's trash problems tomorrow. They simply showed that their "lightweight highlighter" works well enough to be a good starting point for future tools.

Summary

The Peacemaker team built a simple, efficient, and honest tool to help computers understand Italian waste management documents. They showed that even with limited resources, you can build a system that finds important terms reliably, acting as a solid foundation for future, more advanced tools.

Technical Summary: Peacemaker at ATE-IT

Problem Statement
The paper addresses the challenge of Automatic Term Extraction (ATE) within the specific domain of municipal waste management in Italian. While recent advancements in Large Language Models (LLMs) and Transformer architectures have shown promise in few-shot ATE scenarios, accurate labeling remains difficult due to limited annotated training data and the complexity of identifying multi-word expressions (MWEs) that shift across domains. The authors participated in the ATE-IT Shared Task 2026, specifically Subtask A, which requires identifying and extracting domain-specific terms from Italian text related to waste management regulations and procedures. The core challenge lies in developing a method that is both effective and interpretable, capable of running on limited computational resources while maintaining balanced performance across different term frequencies.

Methodology
The proposed system, "Peacemaker," employs a Sequence Labeling Framework based on a fine-tuned Transformer model. The methodology consists of three primary stages:

Data Preprocessing and Alignment:
A custom preprocessing pipeline was developed to convert raw sentence-level term annotations into a format suitable for sequence labeling.
- BIO Tagging: The standard Begin-Inside-Outside (B-I-O) scheme was applied, assigning B-TERM to the first token of a term, I-TERM to subsequent tokens, and O to non-term tokens.
- Token-Term Alignment: To resolve ambiguities between character-level terms and subword tokens, the authors utilized offset mappings from the tokenizer. A masking mechanism was implemented to ensure that each character position maps to only one term, preventing overlapping labels.
- Label Encoding: Symbolic labels were converted to numerical IDs (O=0, B-TERM=1, I-TERM=2), with special tokens and padding masked with -100 to be ignored by the loss function.
Model Architecture:
The core of the system utilizes the dbmdz/bert-base-italian-cased model, a pre-trained cased BERT variant trained on an extensive Italian corpus. A linear layer with a Softmax activation function was appended to the backbone for token classification. The model was fine-tuned using the AdamW optimizer (learning rate $2 \times 10^{-5}$ , weight decay 0.01) over 5 epochs with a batch size of 16.
Evaluation Metrics:
Performance was assessed using two distinct perspectives:
- Type-level Metrics: Evaluating the system's ability to identify unique domain terms regardless of frequency (Precision, Recall, F1 based on unique term counts).
- Micro-level Metrics: Evaluating performance on a sentence-by-sentence basis by aggregating True Positives, False Positives, and False Negatives across the entire dataset.

Key Contributions
The authors highlight three specific contributions to the field of ATE:

Robust Data Alignment: The implementation of a custom pipeline featuring token-level semantic alignment (BIO tagging) specifically tailored for the Italian language and the ATE-IT task, effectively resolving token boundary ambiguities.
High-Fidelity Feature Extraction: The utilization of a context-aware Transformer architecture (dbmdz/bert-base-italian-cased) optimized for token-level predictions within the Italian regulatory domain.
Reproducibility: The explicit pinning of critical Python dependencies (e.g., transformers, accelerate) to ensure environment stability and reproducibility across different runtime environments.

Results
The system was evaluated against a baseline created using a closed-source LLM (gemini-2.5-flash) in zero-shot mode and compared against 8 other participating teams.

Baseline Performance: The zero-shot baseline achieved a Type F1 of 0.469 and a Micro F1 of 0.526.
Peacemaker Performance: The proposed method achieved a Type Precision of 0.430, Type Recall of 0.455, and Type F1 of 0.442. On the micro-level, it achieved a Precision of 0.497, Recall of 0.476, and F1 of 0.486.
Ranking: The team finished 7th out of 9 participating teams.
Observations: The results indicate a balanced performance between Precision and Recall at both type and micro levels. The system demonstrated consistent extraction capabilities for both frequent and less frequent terms, avoiding a reliance solely on high-frequency term detection.

Significance and Claims
The paper positions its findings as a demonstration of the viability of "low-cost and interpretable" methods for ATE. The authors modestly claim that while their approach did not achieve the highest absolute performance (likely due to limited computational resources compared to teams using larger or more specialized models), it serves as a "good starting point for low-resource models."

The significance of the work lies in its stability and balance; the method produces consistent extraction results without over-optimizing for precision through repetitive cycles. The authors conclude that their system remains highly competitive given the resource constraints, suggesting that future advancements in model expansion can retain interpretability while achieving higher performance. The paper explicitly states that Generative AI tools were used only for brainstorming and drafting, with all methodological decisions and analysis conducted by the authors.