---
language: "en"
tags:
- gpt2
- exbert
- commonsense
- semeval2020
- comve
license: "mit"
datasets:
- ComVE
metrics:
- bleu
widget:
- text: "Chicken can swim in water. <|continue|>"
---

# ComVE-gpt2-medium

## Model description

Finetuned model on Commonsense Validation and Explanation (ComVE) dataset introduced in [SemEval2020 Task4](https://competitions.codalab.org/competitions/21080) using a causal language modeling (CLM) objective.
The model is able to generate a reason why a given natural language statement is against commonsense.

## Intended uses & limitations

You can use the raw model for text generation to generate reasons why natural language statements are against commonsense.

#### How to use

You can use this model directly to generate reasons why the given statement is against commonsense using [`generate.sh`](https://github.com/AliOsm/SemEval2020-Task4-ComVE/tree/master/TaskC-Generation) script.

*Note:* make sure that you are using version `2.4.1` of `transformers` package. Newer versions has some issue in text generation and the model repeats the last token generated again and again.

#### Limitations and bias

The model biased to negate the entered sentence usually instead of producing a factual reason.

## Training data

The model is initialized from the [gpt2-medium](https://github.com/huggingface/transformers/blob/master/model_cards/gpt2-README.md) model and finetuned using [ComVE](https://github.com/wangcunxiang/SemEval2020-Task4-Commonsense-Validation-and-Explanation) dataset which contains 10K against commonsense sentences, each of them is paired with three reference reasons.

## Training procedure

Each natural language statement that against commonsense is concatenated with its reference reason with `<|continue|>` as a separator, then the model finetuned using CLM objective.
The model trained on Nvidia Tesla P100 GPU from Google Colab platform with 5e-5 learning rate, 5 epochs, 128 maximum sequence length and 64 batch size.

<center>
  <img src="https://i.imgur.com/xKbrwBC.png">
</center>

## Eval results

The model achieved fifth place with 16.7153/16.1187 BLEU scores and third place with 1.94 Human Evaluation score on SemEval2020 Task4: Commonsense Validation and Explanation development and testing dataset.

These are some examples generated by the model:
|             Against Commonsense Statement             |               Generated Reason               |
|:-----------------------------------------------------:|:--------------------------------------------:|
| Chicken can swim in water.                            | Chicken can't swim.                          |
| shoes can fly                                         | Shoes are not able to fly.                   |
| Chocolate can be used to make a coffee pot            | Chocolate is not used to make coffee pots.   |
| you can also buy tickets online with an identity card | You can't buy tickets with an identity card. |
| a ball is square and can roll                         | A ball is round and cannot roll.             |
| You can use detergent to dye your hair.               | Detergent is used to wash clothes.           |
| you can eat mercury                                   | mercury is poisonous                         |
| A gardener can follow a suspect                       | gardener is not a police officer             |
| cars can float in the ocean just like a boat          | Cars are too heavy to float in the ocean.    |
| I am going to work so I can lose money.               | Working is not a way to lose money.          |

### BibTeX entry and citation info

```bibtex
@article{fadel2020justers,
  title={JUSTers at SemEval-2020 Task 4: Evaluating Transformer Models Against Commonsense Validation and Explanation},
  author={Fadel, Ali and Al-Ayyoub, Mahmoud and Cambria, Erik},
  year={2020}
}
```

<a href="https://huggingface.co/exbert/?model=aliosm/ComVE-gpt2-medium">
	<img width="300px" src="https://cdn-media.huggingface.co/exbert/button.png">
</a>