app.py

import gradio as gr
import numpy as np
import torch
from transformers import AutoTokenizer, AutoModelForSequenceClassification
from transformers import pipeline

emotion_labels = ['喜び', '悲しみ', '期待', '驚き', '怒り', '信頼', '悲しみ', '嫌悪']
sentiment_labels = ['ポジティブ', 'ニュートラル', 'ネガティブ']

def np_softmax(x):
    x_exp = torch.exp(torch.tensor(x) - torch.max(torch.tensor(x)))
    f_x = x_exp / x_exp.sum()
    return f_x

def emotion_classifier(text):
    model.eval()
    tokens = tokenizer(text, truncation=True, return_tensors="pt")
    tokens.to(model.device)
    preds = model(**tokens)
    prob = np_softmax(preds.logits.cpu().detach().numpy()[0])
    out_dict = {n: p.item() for n, p in zip(emotion_labels, prob)}
    return out_dict

tokenizer = AutoTokenizer.from_pretrained("cl-tohoku/bert-base-japanese-whole-word-masking")
model = AutoModelForSequenceClassification.from_pretrained("jingwora/language-emotion-classification-ja", num_labels=8)

def sentiment_classifier(text):
  clf = pipeline(model="lxyuan/distilbert-base-multilingual-cased-sentiments-student", return_all_scores=True)
  sentiment = clf(text)
  sentiment = {item['label']: item['score'] for item in sentiment[0]}
  sentiment = {sentiment_labels[i]: sentiment[label] for i, label in enumerate(sentiment)} # 日本語ラベル
  return sentiment

examples = [
    ["このお店は本当に素晴らしいです！サービスも料理も満足できるものばかりでした。"],
    ["料理の味が期待外れでした。改善が必要ですね。"],
    ["サービスは普通ですが、特に不満もありません。"],
    ["価格と品質のバランスが取れていると思います。"],
]


demo = gr.Blocks(
    theme="freddyaboulton/dracula_revamped",
)

with demo:
    gr.Markdown(
        """
    # Emotion and Sentiment Classification
    Enter Japanese text and get the emotion probabilities and sentiment probablilities.
    """
    )
    text = gr.Textbox(lines=2)
    with gr.Row():
        gr.Examples(examples=examples, inputs=text)

    b1 = gr.Button("Emotion Classification")
    label1 = gr.Label(num_top_classes=8)
    b1.click(emotion_classifier, inputs=text, outputs=label1)

    b2 = gr.Button("Sentiment Analysis")
    label2 = gr.Label(num_top_classes=3)
    b2.click(sentiment_classifier, inputs=text, outputs=label2)

demo.launch()