README.md

---
language: de
tags:
- generated_from_trainer
datasets:
- Short-Answer-Feedback/saf_micro_job_german
widget:
- text: 'Antwort: Ich gebe mich zu erkennen und zeige das Informationsschreiben vor
    Lösung: Der Jobber soll sich in diesem Fall dem Personal gegenüber zu erkennen
    geben (0.25 P) und das entsprechende Informationsschreiben in der App vorzeigen
    (0.25 P). Zusätzlich muss notiert werden, zu welchem Zeitpunkt (0.25 P) des Jobs
    der Jobber enttarnt wurde. Zentrale Frage ist dabei, ob ein neutrales, unvoreingenommenes
    Verkaufsgespräch stattgefunden hat. Der Job soll mit Erlaubnis der Mitarbeiter
    bis zum Ende durchgeführt (0.25 P) werden. Frage: Frage 1: Wie reagierst du, wenn
    du auf deine Tätigkeit angesprochen wirst?'
base_model: facebook/mbart-large-cc25
---

# mbart-finetuned-saf-micro-job

This model is a fine-tuned version of [facebook/mbart-large-cc25](https://huggingface.co/facebook/mbart-large-cc25) on the [saf_micro_job_german](https://huggingface.co/datasets/Short-Answer-Feedback/saf_micro_job_german) dataset for Short Answer Feedback (SAF), as proposed in [Filighera et al., ACL 2022](https://aclanthology.org/2022.acl-long.587).

## Model description

This model was built on top of [mBART](https://arxiv.org/abs/2001.08210), which is a sequence-to-sequence denoising auto-encoder pre-trained on large-scale monolingual corpora in many languages.

It expects inputs in the following format:
```
Antwort: [answer] Lösung: [reference_answer] Frage: [question]
```

In the example above, `[answer]`, `[reference_answer]` and `[question]` should be replaced by the provided answer, the reference answer and the question to which they refer, respectively.


The outputs are formatted as follows:
```
[verification_feedback] Feedback: [feedback]
```

Hence, the `[verification_feedback]` label will be one of `Correct`, `Partially correct` or `Incorrect`, while `[feedback]` will be the textual feedback generated by the model according to the given answer.

## Intended uses & limitations

This model is intended to be used for Short Answer Feedback generation in the context of micro-job training (as conducted on the crowd-worker platform appJobber). Thus, it is not expected to have particularly good performance on sets of questions and answers out of this scope.

It is important to acknowledge that the model underperforms when a question that was not seen during training is given as input for inference. In particular, it tends to classify most answers as being correct and does not provide relevant feedback in such cases. Nevertheless, this limitation could be partially overcome by extending the dataset with the desired question (and associated answers) and fine-tuning it for a few epochs on the new data.  

## Training and evaluation data

As mentioned previously, the model was trained on the [saf_micro_job_german](https://huggingface.co/datasets/Short-Answer-Feedback/saf_micro_job_german) dataset, which is divided into the following splits.

| Split                 | Number of examples |
| --------------------- | ------------------ |
| train                 | 1226	             |
| validation            | 308	             |
| test_unseen_answers   | 271	             |
| test_unseen_questions | 602                |

Evaluation was performed on the `test_unseen_answers` and `test_unseen_questions` splits.

## Training procedure

The [Trainer API](https://huggingface.co/docs/transformers/main_classes/trainer#transformers.Seq2SeqTrainer) was used to fine-tune the model. The code utilized for pre-processing and training was mostly adapted from the [summarization script](https://github.com/huggingface/transformers/tree/main/examples/pytorch/summarization) made available by HuggingFace.

Training was completed in a little under 1 hour on a GPU on Google Colab.

### Training hyperparameters

The following hyperparameters were utilized during training:
- num_epochs: 10
- optimizer: Adam with betas=(0.9, 0.999) and epsilon=1e-08
- learning_rate: 5e-05
- lr_scheduler_type: linear
- train_batch_size: 1
- gradient_accumulation_steps: 4
- eval_batch_size: 4
- mixed_precision_training: Native AMP
- PyTorch seed: 42

### Framework versions

- Transformers 4.25.1
- Pytorch 1.12.1+cu113
- Datasets 2.7.1
- Tokenizers 0.13.2

## Evaluation results

The generated feedback was evaluated through means of the [SacreBLEU](https://huggingface.co/spaces/evaluate-metric/sacrebleu), [ROUGE-2](https://huggingface.co/spaces/evaluate-metric/rouge), [METEOR](https://huggingface.co/spaces/evaluate-metric/meteor), [BERTScore](https://huggingface.co/spaces/evaluate-metric/bertscore) metrics from HuggingFace, while the [accuracy](https://scikit-learn.org/stable/modules/generated/sklearn.metrics.accuracy_score.html) and [F1](https://scikit-learn.org/stable/modules/generated/sklearn.metrics.f1_score.html) scores from scikit-learn were used for evaluation of the labels.

The following results were achieved.

| Split                 | SacreBLEU | ROUGE-2 | METEOR | BERTScore | Accuracy | Weighted F1 | Macro F1 | 
| --------------------- | :-------: | :-----: | :----: | :-------: | :------: | :---------: | :------: |
| test_unseen_answers   | 39.5	    | 29.8    | 63.3   | 63.1      | 80.1     | 80.3        | 80.7     |
| test_unseen_questions | 0.3       | 0.5     | 33.8   | 31.3      | 48.7     | 46.5        | 40.6     |


The script used to compute these metrics and perform evaluation can be found in the `evaluation.py` file in this repository.

## Usage

The example below shows how the model can be applied to generate feedback to a given answer.

```python
from transformers import AutoModelForSeq2SeqLM, AutoTokenizer

model = AutoModelForSeq2SeqLM.from_pretrained('Short-Answer-Feedback/mbart-finetuned-saf-micro-job')
tokenizer = AutoTokenizer.from_pretrained('Short-Answer-Feedback/mbart-finetuned-saf-micro-job')

example_input = 'Antwort: Ich gebe mich zu erkennen und zeige das Informationsschreiben vor Lösung: Der Jobber soll sich in diesem Fall dem Personal gegenüber zu erkennen geben (0.25 P) und das entsprechende Informationsschreiben in der App vorzeigen (0.25 P). Zusätzlich muss notiert werden, zu welchem Zeitpunkt (0.25 P) des Jobs der Jobber enttarnt wurde. Zentrale Frage ist dabei, ob ein neutrales, unvoreingenommenes Verkaufsgespräch stattgefunden hat. Der Job soll mit Erlaubnis der Mitarbeiter bis zum Ende durchgeführt (0.25 P) werden. Frage: Frage 1: Wie reagierst du, wenn du auf deine Tätigkeit angesprochen wirst?'
inputs = tokenizer(example_input, max_length=256, padding='max_length', truncation=True, return_tensors='pt')

generated_tokens = model.generate(
                inputs['input_ids'],
                attention_mask=inputs['attention_mask'],
                max_length=128
            )
output = tokenizer.batch_decode(generated_tokens, skip_special_tokens=True)[0]
```

The output produced by the model then looks as follows:

```
Partially correct Feedback: Sollte das Personal dies gestatten, kannst du den Check auch gerne noch abschließen. Bitte halte nur in fest, wann genau du auf deine Tätigkeit angesprochen wurdest.
```

## Related Work

[Filighera et al., ACL 2022](https://aclanthology.org/2022.acl-long.587) trained a [mT5 model](https://huggingface.co/docs/transformers/model_doc/mt5) on this dataset, providing a baseline for SAF generation. The entire code used to define and train the model can be found on [GitHub](https://github.com/SebOchs/SAF).