File size: 2,781 Bytes
5a705e8
 
 
 
 
 
 
 
0f4d23d
 
 
 
 
5a705e8
 
 
 
 
6bea83a
 
 
5a705e8
 
9c96d0b
5a705e8
 
 
 
 
 
 
 
 
9c96d0b
 
 
5a705e8
 
 
9c96d0b
5a705e8
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
---
tags:
- generated_from_trainer
metrics:
- accuracy
- precision
- recall
- f1
model_creators:
- Jordan Painter, Diptesh Kanojia
widget:
- text: wow, i mean who would have thought
base_model: vinai/bertweet-base
model-index:
- name: bertweet-base-finetuned-SARC-combined-DS
  results: []
---

# Utilising Weak Supervision to Create S3D: A Sarcasm Annotated Dataset
This is the repository for the S3D dataset published at EMNLP 2022. The dataset can help build sarcasm detection models.

# bertweet-base-finetuned-SARC-combined-DS

This model is a fine-tuned version of [vinai/bertweet-base](https://huggingface.co./vinai/bertweet-base) on our combined sarcasm dataset.
It achieves the following results on the evaluation set:
- Loss: 1.4624
- Accuracy: 0.7611
- Precision: 0.7611
- Recall: 0.7611
- F1: 0.7611

## Model description

The given description for BERTweet by VinAI is as follows: <br>
BERTweet is the first public large-scale language model pre-trained for English Tweets. BERTweet is trained based on the RoBERTa pre-training procedure. The corpus used to pre-train BERTweet consists of 850M English Tweets (16B word tokens ~ 80GB), containing 845M Tweets streamed from 01/2012 to 08/2019 and 5M Tweets related to the COVID-19 pandemic.
<br>

## Training and evaluation data

More information neededThis [vinai/bertweet-base](https://huggingface.co./vinai/bertweet-base) model was finetuned on our combined sarcasm dataset. This dataset was created to aid the building of sarcasm detection models

## Training procedure

### Training hyperparameters

The following hyperparameters were used during training:
- learning_rate: 1e-05
- train_batch_size: 64
- eval_batch_size: 16
- seed: 43
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- num_epochs: 30

### Training results

| Training Loss | Epoch | Step   | Validation Loss | Accuracy | Precision | Recall | F1     |
|:-------------:|:-----:|:------:|:---------------:|:--------:|:---------:|:------:|:------:|
| 0.4319        | 4.0   | 44819  | 0.5049          | 0.7790   | 0.7796    | 0.7789 | 0.7789 |
| 0.2835        | 8.0   | 89638  | 0.6475          | 0.7663   | 0.7664    | 0.7663 | 0.7663 |
| 0.1797        | 12.0  | 134457 | 0.8746          | 0.7638   | 0.7639    | 0.7637 | 0.7637 |
| 0.1219        | 16.0  | 179276 | 1.0595          | 0.7585   | 0.7597    | 0.7587 | 0.7583 |
| 0.0905        | 20.0  | 224095 | 1.2115          | 0.7611   | 0.7612    | 0.7612 | 0.7611 |
| 0.0728        | 24.0  | 268914 | 1.3644          | 0.7628   | 0.7629    | 0.7627 | 0.7627 |
| 0.0612        | 28.0  | 313733 | 1.4624          | 0.7611   | 0.7611    | 0.7611 | 0.7611 |


### Framework versions

- Transformers 4.20.1
- Pytorch 1.10.1+cu111
- Datasets 2.3.2
- Tokenizers 0.12.1