| # Dialog-KoELECTRA | |
| Github : [https://github.com/skplanet/Dialog-KoELECTRA](https://github.com/skplanet/Dialog-KoELECTRA) | |
| ## Introduction | |
| **Dialog-KoELECTRA** is a language model specialized for dialogue. It was trained with 22GB colloquial and written style Korean text data. Dialog-ELECTRA model is made based on the [ELECTRA](https://openreview.net/pdf?id=r1xMH1BtvB) model. ELECTRA is a method for self-supervised language representation learning. It can be used to pre-train transformer networks using relatively little compute. ELECTRA models are trained to distinguish "real" input tokens vs "fake" input tokens generated by another neural network, similar to the discriminator of a [GAN](https://arxiv.org/pdf/1406.2661.pdf). At small scale, ELECTRA achieves strong results even when trained on a single GPU. | |
| <br> | |
| ## Released Models | |
| We are initially releasing small version pre-trained model. | |
| The model was trained on Korean text. We hope to release other models, such as base/large models, in the future. | |
| | Model | Layers | Hidden Size | Params | Max<br/>Seq Len | Learning<br/>Rate | Batch Size | Train Steps | | |
| | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | | |
| | Dialog-KoELECTRA-Small | 12 | 256 | 14M | 128 | 1e-4 | 512 | 700K | | |
| <br> | |
| ## Model Performance | |
| Dialog-KoELECTRA shows strong performance in conversational downstream tasks. | |
| | | **NSMC**<br/>(acc) | **Question Pair**<br/>(acc) | **Korean-Hate-Speech**<br/>(F1) | **Naver NER**<br/>(F1) | **KorNLI**<br/>(acc) | **KorSTS**<br/>(spearman) | | |
| | :--------------------- | :----------------: | :--------------------: | :----------------: | :------------------: | :-----------------------: | :-------------------------: | | |
| | DistilKoBERT | 88.60 | 92.48 | 60.72 | 84.65 | 72.00 | 72.59 | | |
| | **Dialog-KoELECTRA-Small** | **90.01** | **94.99** | **68.26** | **85.51** | **78.54** | **78.96** | | |
| <br> | |
| ## Train Data | |
| <table class="tg"> | |
| <thead> | |
| <tr> | |
| <th class="tg-c3ow"></th> | |
| <th class="tg-c3ow">corpus name</th> | |
| <th class="tg-c3ow">size</th> | |
| </tr> | |
| </thead> | |
| <tbody> | |
| <tr> | |
| <td class="tg-c3ow" rowspan="4">dialog</td> | |
| <td class="tg-0pky"><a href="https://aihub.or.kr/aidata/85" target="_blank" rel="noopener noreferrer">Aihub Korean dialog corpus</a></td> | |
| <td class="tg-c3ow" rowspan="4">7GB</td> | |
| </tr> | |
| <tr> | |
| <td class="tg-0pky"><a href="https://corpus.korean.go.kr/" target="_blank" rel="noopener noreferrer">NIKL Spoken corpus</a></td> | |
| </tr> | |
| <tr> | |
| <td class="tg-0pky"><a href="https://github.com/songys/Chatbot_data" target="_blank" rel="noopener noreferrer">Korean chatbot data</a></td> | |
| </tr> | |
| <tr> | |
| <td class="tg-0pky"><a href="https://github.com/Beomi/KcBERT" target="_blank" rel="noopener noreferrer">KcBERT</a></td> | |
| </tr> | |
| <tr> | |
| <td class="tg-c3ow" rowspan="2">written</td> | |
| <td class="tg-0pky"><a href="https://corpus.korean.go.kr/" target="_blank" rel="noopener noreferrer">NIKL Newspaper corpus</a></td> | |
| <td class="tg-c3ow" rowspan="2">15GB</td> | |
| </tr> | |
| <tr> | |
| <td class="tg-0pky"><a href="https://github.com/lovit/namuwikitext" target="_blank" rel="noopener noreferrer">namuwikitext</a></td> | |
| </tr> | |
| </tbody> | |
| </table> | |
| <br> | |
| ## Vocabulary | |
| We applied morpheme analysis using [huggingface_konlpy](https://github.com/lovit/huggingface_konlpy) when creating a vocabulary dictionary. | |
| As a result of the experiment, it showed better performance than a vocabulary dictionary created without applying morpheme analysis. | |
| <table> | |
| <thead> | |
| <tr> | |
| <th>vocabulary size</th> | |
| <th>unused token size</th> | |
| <th>limit alphabet</th> | |
| <th>min frequency</th> | |
| </tr> | |
| </thead> | |
| <tbody> | |
| <tr> | |
| <td>40,000</td> | |
| <td>500</td> | |
| <td>6,000</td> | |
| <td>3</td> | |
| </tr> | |
| </tbody> | |
| </table> | |
| <br> | |