Multilingual Cpv Sector Classifier

🚀 Multilingual CPV Sector Classifier

This model is a fine - tuned version of [bert - base - multilingual - cased](https://huggingface.co/bert - base - multilingual - cased) for classifying procurement descriptions into sector classes, achieving an F1 Score of 0.686 on the evaluation set.

🚀 Quick Start

This model is a fine - tuned version of [bert - base - multilingual - cased](https://huggingface.co/bert - base - multilingual - cased) on the Tenders Economic Daily Public Procurement Data. It achieves the following results on the evaluation set:

• F1 Score: 0.686

✨ Features

• Multilingual Support: The model can handle procurement descriptions written in any of [104 languages](https://github.com/google - research/bert/blob/master/multilingual.md#list - of - languages).
• Sector Classification: It classifies procurement descriptions into 45 sector classes represented by CPV(Common Procurement Vocabulary) code descriptions.

📚 Documentation

Model description

The model takes procurement descriptions written in any of [104 languages](https://github.com/google - research/bert/blob/master/multilingual.md#list - of - languages) and classifies them into 45 sector classes represented by CPV(Common Procurement Vocabulary) code descriptions as listed below.

Intended uses & limitations

• Input description should be written in any of [the 104 languages](https://github.com/google - research/bert/blob/master/multilingual.md#list - of - languages) that MBERT supports.
• The model is just evaluated in 22 languages. Thus there is no information about the performances in other languages.
• The domain is also restricted by the awarded procurement notice descriptions in European Union. Evaluating on whole document texts might change the performance.

Training and evaluation data

• The whole data consists of 744,360 rows. Shuffled and split into train and validation sets by using 80%/20% manner.
• Each description represents a unique contract notice description awarded between 2011 and 2018.
• Both training and validation data have contract notice descriptions written in 22 European Languages. (Malta and Irish are extracted due to scarcity compared to whole data)

Training procedure

The training procedure has been completed on Google Cloud V3 - 8 TPUs. Thanks Google for giving the access to Cloud TPUs

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 2e - 05
• num_epochs: 3
• gradient_accumulation_steps: 8
• batch_size_per_device: 4
• total_train_batch_size: 32

Training results

📄 License

This model is licensed under the Apache - 2.0 license.