---
tags:
- generated_from_trainer
metrics:
- precision
- recall
- f1
- accuracy
model-index:
- name: bert-base-spanish-wwm-uncased-finetuned-NER-medical
  results: []
widget:
- text: "El útero o matriz es el lugar donde se desarrolla el bebé cuando una mujer está embarazada."
- text: "El síndrome de dolor regional complejo es un trastorno de dolor crónico."

---
# bert-base-spanish-wwm-uncased-finetuned-NER-medical

This model is a fine-tuned version of [dccuchile/bert-base-spanish-wwm-uncased](https://huggingface.co/dccuchile/bert-base-spanish-wwm-uncased) on an adaptation of [eHealth-KD Challenge 2020 dataset](https://knowledge-learning.github.io/ehealthkd-2020/), filtered only for the task of NER. The dataset annotations for NER are ['Concept', 'Action', 'Predicate', 'Reference']. 

Before the training process, the dataset had processed to label it with the BIO annotation format. Some cleaning and adaptations were needed, for example, to work with overlapped entities. 


It achieves the following results on the evaluation set:
- Loss: 0.6433
- Precision: 0.8297
- Recall: 0.8367
- F1: 0.8332
- Accuracy: 0.8876

## Model description

A BERT adaptation for Spanish medical NER. This type of models can be part of NLP pipelines created, for example, to analyse clinical documents, automatic labelling of clinical documents following standard classifications such as CIE-10 o SNOMED, etc.

## Training and evaluation data

The adapted dataset has this structure:

- Training: 800 labelled sentences
- Development: 199 labelled sentences
- Testing: 100 labelled sentences

## Training procedure

The chapter  [“Token classification”]( https://huggingface.co/course/chapter7/2) in the Hugging Face online course was used as starting point for the training process. We made some adaptions because our dataset follows a slightly different structure. Moreover, a conversion between string labels and integers labels was needed.

### Training hyperparameters

The following hyperparameters were used during training:
- learning_rate: 2e-05
- train_batch_size: 16
- eval_batch_size: 16
- seed: 42
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- num_epochs: 12

### Training results

| Training Loss | Epoch | Step | Validation Loss | Precision | Recall | F1     | Accuracy |
|:-------------:|:-----:|:----:|:---------------:|:---------:|:------:|:------:|:--------:|
| 0.1139        | 1.0   | 50   | 0.3932          | 0.8671    | 0.8378 | 0.8522 | 0.9004   |
| 0.074         | 2.0   | 100  | 0.4334          | 0.8682    | 0.8367 | 0.8522 | 0.9004   |
| 0.0564        | 3.0   | 150  | 0.4498          | 0.8654    | 0.8353 | 0.8501 | 0.8993   |
| 0.0431        | 4.0   | 200  | 0.4683          | 0.8629    | 0.8425 | 0.8526 | 0.8985   |
| 0.0328        | 5.0   | 250  | 0.4850          | 0.8508    | 0.8454 | 0.8481 | 0.8964   |
| 0.027         | 6.0   | 300  | 0.4983          | 0.8608    | 0.8432 | 0.8519 | 0.8988   |
| 0.0253        | 7.0   | 350  | 0.5334          | 0.8618    | 0.8457 | 0.8537 | 0.9004   |
| 0.0242        | 8.0   | 400  | 0.5546          | 0.8636    | 0.8450 | 0.8542 | 0.9009   |
| 0.0233        | 9.0   | 450  | 0.5507          | 0.8543    | 0.8436 | 0.8489 | 0.8961   |
| 0.0203        | 10.0  | 500  | 0.5410          | 0.8605    | 0.8432 | 0.8518 | 0.9001   |
| 0.0179        | 11.0  | 550  | 0.5547          | 0.8603    | 0.8507 | 0.8555 | 0.9006   |
| 0.0149        | 12.0  | 600  | 0.5568          | 0.8616    | 0.8446 | 0.8531 | 0.9012   |


### Framework versions

- Transformers 4.17.0
- Pytorch 1.10.0+cu111
- Datasets 2.0.0
- Tokenizers 0.11.6