00ddya3 / k-finbert

Financial Sentiment Anlaysis

K-FinBERT: Korean Financial Sentiment Analysis with BERT

This repository is fine-tuning FinBERT with Korean financial sentence dataset from AIHub. For detailed information about FinBERT, see the FinBERT.

Also, you can view learning curves in the wandb

Prepare Datasets

Finance Sentiment Corpus
This dataset translate 'Finance Phrase Bank' from English to Korean

Download
Finance Sentiment Corpus

AIHub
The Korean sentence dataset labeled from news, magazines, broadcast scripts, blogs, and books format with various categories, like history, society, finance, IT sience, and etc. Sentence is labeld by

• type: conversation/fact/inference/predict
• certainty: certain/uncertain
• temporality: past/present/future
• sentiment: positive/negative/neutral

We only use the finance news dataset for training.

Download
Download the '문장 유형(추론, 예측 등) 판단 데이터' dataset from AIHub. As the dataset is split compressed, a program that supports split decompression (e.g. Bandizip, 7-Zip) is recommended.

Preprocess
Dataset should be preprocessed to {'ID': int, 'Text': str, 'Label': str} form and saved as csv with delimiter='\t'.

python ./scripts/data_utils.py --data_path "your_directory/finance_data.csv"
python ./scripts/data_utils.py --data_path "your_directory/TL_뉴스_금융.zip" # not zip file but folder
python ./scripts/data_utils.py --data_path "your_directory/VL_뉴스_금융.zip"

Installing

Repository
Clone this repository

git clone https://github.com/gynchoi/finBERT.git

Environments
Create the Conda environment

conda env create -f environment.yml
conda activate finbert

Pre-trained model Checkpoints
Download the original FinBERT checkpoint from HuggingFace/FinBERT.

mkdir models/sentiment
cd ./models/sentiment/

git lfs install
git clone https://huggingface.co/ProsusAI/finbert

If you got git: 'lfs' is not a git command. See 'git --help error, install git-lfs in your terminal first.

sudo apt install git-lfs

Minor Modifications

You can neglect belows, if you clone this repository.

Tokenizer
For utilizing the Korean dataset, tokenizer is changed from 'bert-base-uncased' to 'monologg/kobert'. To use KoBERT tokenizer

1. Copy tokenization_kobert.py to ./finbert/ folder
2. Download sentencepiece package, unless you may get UnboundLocalError: local variable 'spm' referenced before assignment error

   pip install sentencepiecc

3. Modify the './finbert/finbert.py' code

   from finbert.tokenization_kobert import KoBertTokenizer
   
   # tokenizer = AutoTokenizer.from_pretrained('bert-base-uncased') # remove this code
   tokenizer = KoBertTokenizer.from_pretrained('monologg/kobert')

Encoding
When open Korean dataset, encoding is needed. Change the './finbert/utils' as below

# with open(input_file, "r") as f
with open(input_file, "r",  encoding='utf-8') as f

Trainer
Since original FinBERT trainer code is presented with jupyter notebook, we rewrite the './notebooks/finbert_training.ipynb' to python format

1. Joining paths in 'finbert.py' with OS package

# self.config.model_dir / ('temporary' + str(best_model)
import os

os.path.join(self.config.model_dir, ('temporary' + str(best_model)))

Test
From predict.py, we need to download and import nltk

import nltk
nltk.download('punkt')

Errors Report

You can view expected errors and some solutions in ERRORS.md