Giyaseddin/distilroberta-base-finetuned-short-answer-assessment

DistilRoBERTa base model for Short Question Answer Assessment

Model description

The pre-trained model is a distilled version of the RoBERTa-base model. It follows the same training procedure as DistilBERT. The code for the distillation process can be found here. This model is case-sensitive: it makes a difference between english and English.

The model has 6 layers, 768 dimension and 12 heads, totalizing 82M parameters (compared to 125M parameters for RoBERTa-base). On average DistilRoBERTa is twice as fast as Roberta-base.

We encourage to check RoBERTa-base model to know more about usage, limitations and potential biases.

This is a classification model that solves Short Question Answer Assessment task, finetuned pretrained DistilRoBERTa model on Question Answer Assessment dataset

Intended uses & limitations

This can only be used for the kind of questions and answers provided by that are similar to the ones in the dataset of Banjade et al..

How to use

You can use this model directly with a :

>>> from transformers import pipeline
>>> classifier = pipeline("text-classification", model="Giyaseddin/distilroberta-base-finetuned-short-answer-assessment", return_all_scores=True)
>>> context = "To rescue a child who has fallen down a well, rescue workers fasten him to a rope, the other end of which is then reeled in by a machine. The rope pulls the child straight upward at steady speed."
>>> question = "How does the amount of tension in the rope compare to the downward force of gravity acting on the child?"
>>> ref_answer = "Since the child is being raised straight upward at a constant speed, the net force on the child is zero and all the forces balance. That means that the tension in the rope balances the downward force of gravity."
>>> student_answer = "The tension force is higher than the force of gravity."
>>> 
>>> body = " [SEP] ".join([context, question, ref_answer, student_answer])
>>> raw_results = classifier([body])
>>> raw_results
[[{'label': 'LABEL_0', 'score': 0.0004029414849355817},
  {'label': 'LABEL_1', 'score': 0.0005476847873069346},
  {'label': 'LABEL_2', 'score': 0.998059093952179},
  {'label': 'LABEL_3', 'score': 0.0009902542224153876}]]
>>> _LABELS_ID2NAME = {0: "correct", 1: "correct_but_incomplete", 2: "contradictory", 3: "incorrect"}
>>> results = []
>>> for result in raw_results:
        for score in result:
            results.append([
                {_LABELS_ID2NAME[int(score["label"][-1:])]: "%.2f" % score["score"]}
            ])

>>> results
[[{'correct': '0.00'}],
 [{'correct_but_incomplete': '0.00'}],
 [{'contradictory': '1.00'}],
 [{'incorrect': '0.00'}]]

Limitations and bias

Even if the training data used for this model could be characterized as fairly neutral, this model can have biased predictions. It also inherits some of the bias of its teacher model.

This bias will also affect all fine-tuned versions of this model.

Also one of the limiations of this model is the length, longer sequences would lead to wrong predictions, due to the pre-processing phase (after concatentating the input sequences, the important student answer might be pruned!)

Pre-training data

Training data

The RoBERTa model was pretrained on the reunion of five datasets:

• BookCorpus, a dataset consisting of 11,038 unpublished books;
• English Wikipedia (excluding lists, tables and headers) ;
• CC-News, a dataset containing 63 millions English news articles crawled between September 2016 and February 2019.
• OpenWebText, an opensource recreation of the WebText dataset used to train GPT-2,
• Stories a dataset containing a subset of CommonCrawl data filtered to match the story-like style of Winograd schemas.

Together theses datasets weight 160GB of text.

Fine-tuning data

The annotated dataset consists of 900 students’ short constructed answers and their correctness in the given context. Four qualitative levels of correctness are defined, correct, correct-but-incomplete, contradictory and Incorrect.

Training procedure

Preprocessing

In the preprocessing phase, the following parts are concatenated: question context, question, reference_answer, and student_answer using the separator [SEP]. This makes the full text as:

[CLS] Context Sentence [SEP] Question Sentence [SEP] Reference Answer Sentence [SEP] Student Answer Sentence [CLS]

The data are splitted according to the following ratio:

• Training set 80%.
• Test set 20%.

Lables are mapped as: {0: "correct", 1: "correct_but_incomplete", 2: "contradictory", 3: "incorrect"}

Fine-tuning

The model was finetuned on GeForce GTX 960M for 20 minuts. The parameters are:

Parameter	Value
Learning rate	5e-5
Weight decay	0.01
Training batch size	8
Epochs	4

Here is the scores during the training:

Epoch	Training Loss	Validation Loss	Accuracy	F1	Precision	Recall
1	No log	0.773334	0.713706	0.711398	0.746059	0.713706
2	1.069200	0.404932	0.885279	0.884592	0.886699	0.885279
3	0.473700	0.247099	0.931980	0.931675	0.933794	0.931980
3	0.228000	0.205577	0.954315	0.954210	0.955258	0.954315

Evaluation results

When fine-tuned on downstream task of Question Answer Assessment 4 class classification, this model achieved the following results: (scores are rounded to 2 floating points)

	precision	recall	f1-score	support
correct	0.933	0.992	0.962	366
correct_but_incomplete	0.976	0.934	0.954	257
contradictory	0.938	0.929	0.933	113
incorrect	0.975	0.932	0.953	249
accuracy	-	-	0.954	985
macro avg	0.955	0.947	0.950	985
weighted avg	0.955	0.954	0.954	985

Confusion matrix:

Actual \ Predicted	correct	correct_but_incomplete	contradictory	incorrect
correct	363	3	0	0
correct_but_incomplete	14	240	0	3
contradictory	5	0	105	3
incorrect	7	3	7	232

The AUC score is: 'micro'= 0.9695 and 'macro': 0.9650