Sentiment Analysis and Text Classification
5.1 Dictionary-based Sentiment Analysis
| Tool | Weekly DL | Total DL | Stars | Description |
|---|---|---|---|---|
| oseti | 169 | 178k | 99 | Based on Japanese sentiment polarity dictionary |
| pymlask | 107 | 68k | 119 | Emotion analysis based on ML-Ask |
oseti is a sentiment analysis tool based on the sentiment polarity dictionary of the Iwanami Dictionary of Japanese, and despite being rule-based, it allows easy positive/negative determination of Japanese text.
pymlask is a tool capable of determining multiple emotion categories such as "joy," "sadness," and "anger," based on the emotion word dictionary ML-Ask.
5.2 Machine Learning-based Sentiment Analysis
| Tool | Weekly DL | Total DL | Stars | Description |
|---|---|---|---|---|
| asari | 91 | 84k | 153 | Japanese sentiment analyzer |
| negapoji | - | - | 152 | Negative/positive determination |
asari has the highest name recognition in the sentiment analysis category at Stars: 153. It includes a built-in machine learning model and enables sentiment analysis through a simple API.
5.3 BERT Fine-tuning Approaches
In recent years, the approach of fine-tuning pretrained BERT models for sentiment analysis tasks has been becoming mainstream.
- Japanese-BERT-Sentiment-Analyzer: A sentiment analysis server using FastAPI and BERT
- japanese_text_classification: Surveys diverse DNN approaches including MLP, CNN, RNN, and BERT
The JGLUE benchmark includes tasks related to sentiment analysis, enabling standard evaluation of BERT-based approaches.
Sentiment Analysis Models on Hugging Face
In the text-classification category on Hugging Face, multiple fine-tuned models for Japanese sentiment analysis have been published.
| Model | HF DL | Likes | Description |
|---|---|---|---|
| bert-finetuned-japanese-sentiment (christian-phu) | 329k | 16 | Trained on Amazon reviews; ~81% accuracy |
| japanese-sentiment-analysis (jarvisx17) | 4k | 15 | Trained on chABSA dataset |
| bert-base-japanese-v2-wrime-fine-tune (patrickramos) | 1k | 6 | WRIME 8-emotion intensity prediction |
bert-finetuned-japanese-sentiment shows the largest usage in the text-classification category at 329k weekly downloads. It fine-tunes cl-tohoku/bert-base-japanese-v2 on Amazon product reviews, reporting roughly 81% accuracy and an F1 score of 0.73.
bert-base-japanese-v2-wrime-fine-tune was fine-tuned on the WRIME dataset and predicts intensity scores (0-4) for eight emotions — joy, sadness, anticipation, surprise, anger, fear, disgust, and trust — from both the writer's and the reader's perspective.