Multi-Language Named Entity Recognition (NER) Analysis

This repository contains a comprehensive Multi-Language NER analysis project utilizing XLM-RoBERTa for Named Entity Recognition tasks across multiple languages. The project demonstrates cross-lingual transfer capabilities and includes both fine-tuned and zero-shot models. It features a Gradio-based application for real-time NER predictions, making it interactive and user-friendly.

Table of Contents

• Project Overview
• Dataset
• Understanding the Dataset
• Data Preprocessing
• Tokenizer Comparison: BERT vs XLM-R
• Tokenizer for NER Analysis
• Model Metrics
• Model Training
• Cross-Lingual Transfer
• Zero-Shot vs Fine-Tuned Model
• Gradio Application
• Installation
• Usage
• Uploading to Hugging Face
• Contributing
• License

Project Overview

This project investigates the application of a multi-language transformer model, specifically XLM-RoBERTa, for Named Entity Recognition (NER) across languages. By fine-tuning on a German dataset and transferring the model to other languages such as French, Italian, and English, the project explores cross-lingual transfer in NER tasks.

Dataset

The primary dataset for this project is the XTREME dataset, commonly used for benchmarking multi-lingual models. We specifically use the German, French, Italian, and English subsets for NER analysis. The dataset includes tokens, labels, and language identifiers.

Understanding the Dataset

Each dataset entry consists of:

• Tokens: Individual words or subwords for token-level classification.
• Labels: Corresponding entity labels (e.g., Person, Organization, Location).
• Language: The language in which the data is written.

This section of the project inspects the distribution of labels and languages, ensuring a balanced dataset representative of multiple languages.

Data Preprocessing

Preprocessing includes:

1. Converting entity labels into numerical representations.
2. Tokenizing text for model compatibility.
3. Splitting data into training, validation, and test sets based on language proportions.

Tokenizer Comparison: BERT vs XLM-R

Comparing BERT and XLM-RoBERTa tokenizers reveals differences in how tokens are processed:

• BERT: Uses special tokens [CLS] and [SEP].
• XLM-R: Uses language-specific encoding and subword segmentation with an underscore (_).

Tokenizer for NER Analysis

The XLM-RoBERTa tokenizer is configured specifically for token classification in NER, allowing robust handling of multiple languages and entity labels.

Model Metrics

Standard NER metrics are used for evaluation:

• Precision, Recall, F1 Score: To evaluate entity-level performance across languages.
• Per-language metrics: To assess cross-lingual transfer effectiveness.

Model Training

The XLM-RoBERTa model is fine-tuned on German NER data. Key steps include:

• Setting up an NER-friendly architecture with token classification layers.
• Training with early stopping and monitoring loss for optimal performance.

Cross-Lingual Transfer

After fine-tuning on German data, the model is tested on French, Italian, and English to observe cross-lingual transfer capabilities. This section analyzes performance in zero-shot settings.

Zero-Shot vs Fine-Tuned Model

We assess the difference between zero-shot (direct application without fine-tuning) and fine-tuned models. Fine-tuning improves accuracy significantly across languages, demonstrating the benefits of transfer learning in NER tasks.

Gradio Application

A Gradio-based web application allows users to interact with the NER model in real time. Users can input text in different languages, and the application will highlight recognized entities.

Example Usage:

• Enter text in one of the supported languages.
• The application will display the entities identified, along with their labels.

Installation

To run this project locally: Additional dependencies for Gradio: bash pip install gradio

Install the Hugging Face Transformers library: bash pip install transformers

Usage

1. Run the Gradio application:

   python gradio_app.py

2. Model Fine-tuning and Evaluation: Run the provided notebook for full training and evaluation steps.

Uploading to Hugging Face

Once fine-tuned, the model can be uploaded to Hugging Face for easy sharing:

1. Authenticate:

   huggingface-cli login

2. Upload the model:

   from transformers import AutoModelForTokenClassification
   
   model = AutoModelForTokenClassification.from_pretrained("path_to_your_fine_tuned_model")
   model.push_to_hub("your_hf_username/your_model_name")

my demo is

https://huggingface.co/spaces/RidaDogrul/NER-Demo

Contributing

Contributions are welcome! Please fork this repository, create a new branch for your feature, and submit a pull request.

License

This project is licensed under the MIT License.