German Emotions

🚀 German-Emotions

This model is a German translation of arpanghoshal/EmoRoBERTa. It enables the classification of 28 emotions in German transcripts, offering valuable insights for emotion analysis in German text.

🚀 Quick Start

This model is a German adaptation of arpanghoshal/EmoRoBERTa. We utilized the go_emotions dataset, translated it into German, and fine - tuned the FacebookAI/xlm - roberta - base model. It can classify 28 emotions in German transcripts, including 'admiration', 'amusement', 'anger', 'annoyance', 'approval', 'caring', 'confusion', 'curiosity', 'desire', 'disappointment', 'disapproval', 'disgust', 'embarrassment', 'excitement', 'fear', 'gratitude', 'grief', 'joy', 'love', 'nervousness', 'optimism', 'pride', 'realization', 'relief', 'remorse', 'sadness', 'surprise', 'neutral'. For detailed information, refer to our published paper at the bottom.

✨ Features

• Capable of classifying 28 different emotions in German transcripts.
• Fine - tuned on a translated German version of the go_emotions dataset.

📦 Installation

To use this model, you need to install the necessary libraries. You can use the following commands:

pip install transformers[torch]
pip install pandas, transformers, numpy, tqdm, openpyxl

💻 Usage Examples

Basic Usage

# pip install transformers[torch]
# pip install pandas, transformers, numpy, tqdm, openpyxl
import pandas as pd
import torch
from transformers import AutoTokenizer, AutoModelForSequenceClassification, Trainer
import numpy as np
from tqdm import tqdm
import time
import os
from transformers import DataCollatorWithPadding
import json

# create base path and input and output path for the model folder and the file folder
base_path = "/share/users/staff/c/clalk/Emotionen"
model_path = os.path.join(base_path, 'Modell')
file_path = os.path.join(base_path, 'Datensatz')

MODEL = "ChrisLalk/German-Emotions"
tokenizer = AutoTokenizer.from_pretrained(MODEL, do_lower_case=False)
model = AutoModelForSequenceClassification.from_pretrained(
    model_path,
    from_tf=False,
    from_flax=False,
    trust_remote_code=False,
    num_labels=28,
    ignore_mismatched_sizes=True
)
data_collator = DataCollatorWithPadding(tokenizer=tokenizer)

# Path to the file
os.chdir(file_path)
df_full = pd.read_excel("speech_turns_pat.xlsx", index_col=None)

if 'Unnamed: 0' in df_full.columns:
    df_full = df_full.drop(columns=['Unnamed: 0'])

df_full.reset_index(drop=True, inplace=True)

# Tokenization and inference function
def infer_texts(texts):
    tokenized_texts = tokenizer(texts, return_tensors="pt", padding=True, truncation=True)
    class SimpleDataset:
        def __init__(self, tokenized_texts):
            self.tokenized_texts = tokenized_texts
        def __len__(self):
            return len(self.tokenized_texts["input_ids"])
        def __getitem__(self, idx):
            return {k: v[idx] for k, v in self.tokenized_texts.items()}
    test_dataset = SimpleDataset(tokenized_texts)
    trainer = Trainer(model=model, data_collator=data_collator)
    predictions = trainer.predict(test_dataset)
    sigmoid = torch.nn.Sigmoid()
    probs = sigmoid(torch.Tensor(predictions.predictions))
    return np.round(np.array(probs), 3).tolist()

start_time = time.time()
df = df_full

# Save results in a dict, here the df contains the additional variables File, Class, session, short_id, long_id, Prediction, hscl-11, and srs.
# However, only the "Sentence" column with the text is relevant for the pipeline. 
results = []
for index, row in tqdm(df.iterrows(), total=df.shape[0]):
    patient_texts = row['Patient']
    prob_list = infer_texts(patient_texts)
    results.append({
        "File": row['Class']+"_"+row['session'],
        "Class": row['Class'],
        "session": row['session'],
        "short_id": row["short_id"],
        "long_id": row["long_id"],
        "Sentence": patient_texts,
        "Prediction": prob_list[0],
        "hscl-11": row["Gesamtscore_hscl"],
        "srs": row["srs_ges"],
    })

# Convert results to df
df_results = pd.DataFrame(results)
df_results.to_json("emo_speech_turn_inference.json")

end_time = time.time()
elapsed_time = end_time - start_time
print(f"Elapsed time: {elapsed_time:.2f} seconds")
print(df_results)

emo_df = pd.DataFrame(df_results['Prediction'].tolist(), index=df_results["Class"].index)
col_names = ['admiration', 'amusement', 'anger', 'annoyance', 'approval', 'caring', 'confusion', 'curiosity', 'desire', 'disappointment', 'disapproval', 'disgust', 'embarrassment', 'excitement', 'fear', 'gratitude', 'grief', 'joy', 'love', 'nervousness', 'optimism', 'pride', 'realization', 'relief', 'remorse', 'sadness', 'surprise', 'neutral']
emo_df.columns = col_names
print(emo_df)

📚 Documentation

Model Details

Classification Metrics

📄 License

This model is licensed under the apache - 2.0 license.

Citation

When using our model, please cite the associated peer - reviewed paper:

@article{Lalk2025EmotionDetection, 
  author = {Christopher Lalk and Kim Targan and Tobias Steinbrenner and Jana Schaffrath and Steffen Eberhardt and Brian Schwartz and Antonia Vehlen and Wolfgang Lutz and Julian Rubel}, 
  title = {Employing large language models for emotion detection in psychotherapy transcripts}, 
  journal = {Frontiers in Psychiatry}, 
  volume = {16}, 
  year = {2025}, 
  doi = {10.3389/fpsyt.2025.1504306}
}