This is a script to perform fine-tuning and inference on pre-trained Korean language models using the NSMC dataset.

The Naver Sentiment Movie Corpus (NSMC) consists of 200K annotated movie reviews. This dataset can be used to finetune pre-trained language models for Korean sentiment analysis.

• ratings_train.txt: 150K reviews for training
• ratings_test.txt: 50K reviews held out for testing

You can install the required dependencies from requirements.txt:

pip install -r requirements.txt

Also install the CUDA or CPU version of PyTorch 2.4.0 from here.

You can train and evaluate a model on the NSMC corpus directly from Huggingface using the following command, where model_name links to any Huggingface model.

python train.py --model [model_name]

Training arguments:

After you have trained a model, you can use that model to predict whether a piece of text has a negative or postive sentiment. Provide a .txt file consisting of one text sample per line:

python predict.py --model [model_name] --input-path [path] --output-path [path]

Example output:

document	label	score
재미지다!	1	0.9554
재밋어~~ 천정명 나오니깐 봄♥♥♥♥	1	0.9829
수호천사 케빈 스케이시~ 좀 잘 만들지....	0	0.5127
3류 막장 야쿠자극화,	0	0.9987
난 왜 지루하게만 느껴지지?	0	0.9948

Alternatively, you can use pipeline.py to predict the sentiment of a single word or sentence:

python pipeline.py --model [model_name] --input ["word_or_sentence"]

Example output:

[{ "input": "노잼", "label": "Negative", "score": "0.9987" }]

Below are some pre-trained language models and their performance on NSMC. Each model was fine-tuned on the ratings_train set and evaluated on the ratings_test set. Batch sizes were set to 64 and each model was trained for 3 epochs, starting at a learning rate of 5e-5.