This model was built on top of mBART, 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 domain of the German social law. 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_legal_domain_german dataset, which is divided into the following splits.

Split	Number of examples
train	1596
validation	400
test_unseen_answers	221
test_unseen_questions	275

Evaluation was performed on the test_unseen_answers and test_unseen_questions splits.

Training procedure

The Trainer API was used to fine-tune the model. The code utilized for pre-processing and training was mostly adapted from the summarization script made available by HuggingFace.

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

Training hyperparameters

The following hyperparameters were used during training:

num_epochs: 9
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
seed: 42

Framework versions

Transformers 4.25.1
Pytorch 1.13.0+cu116
Datasets 2.7.1
Tokenizers 0.13.2

Evaluation results

The generated feedback was evaluated through means of the SacreBLEU, ROUGE-2, METEOR, BERTScore metrics from HuggingFace, while the accuracy and F1 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	42.8	43.7	58.2	57.5	81.0	80.1	74.6
test_unseen_questions	3.2	5.0	20.0	14.8	60.7	55.3	55.4

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-legal-domain')
tokenizer = AutoTokenizer.from_pretrained('Short-Answer-Feedback/mbart-finetuned-saf-legal-domain')

example_input = 'Antwort: Wird sich nicht an die Auflagen gehalten (unzureichende Eigenbemühung), droht eine Sperrzeit von 1-2 Wochen. Dadurch wird für die genannte zeit keine Leistung gezahlt, die Anspruchsdauer vermindert sich insgesamt. Bei wichtigen Gründen wird die Sperrzeit nicht verordnet. Lösung: Merkblatt 1 für Arbeitslose, S. 22: Erbringen Sie die Pflichten im Zusammenhang mit den Eigenbemühungen nicht, nicht rechtzeitig oder nicht vollständig, tritt eine Sperrzeit (0,75 p) ein. Merkblatt 1 für Arbeitslose, S. 55: Die Dauer einer Sperrzeit bei unzureichenden Eigenbemühungen beträgt zwei Wochen. (0,25 p). Frage: Mit welcher Folge und welcher Dauer müssen Sie rechnen, wenn Sie Ihre notwendigen Eigenbemühungen nicht rechtzeitig oder nicht vollständig erfüllen?'
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: Es ist richtig, dass Sie mit einer Sperrzeit rechnen müssen, in der Sie keine Leistung bekommen. Die gesetzlich vorgesehene Sperrzeit bei unzureichenden Eigenbemühungen beträgt jedoch zwei Wochen.