This repository contains a powerful multilingual, multi-label emotion classification model. It is fine-tuned from the robust bert-base-multilingual-cased model on the comprehensive multilingual_go_emotions dataset. The model is designed to analyze text and identify a wide spectrum of 27 different emotions, plus a neutral category. Its ability to detect multiple emotions simultaneously makes it highly effective for understanding nuanced text from diverse sources.
Model Name: AnasAlokla/multilingual_go_emotions_V1.2
Architecture: BERT (bert-base-multilingual-cased)
Tasks: Multi-Label Text Classification | Emotion Detection | Sentiment Analysis
🌍 Truly Multilingual: Natively supports 6 major languages, making it ideal for global applications.
🏷️ Multi-Label Classification: Capable of detecting multiple emotions in a single piece of text, capturing complex emotional expressions.
💪 High Performance: Built on bert-base-multilingual-cased, delivering strong results across all supported languages and emotions. See the detailed evaluation metrics.
🔗 Open & Accessible: Comes with a live demo, the full dataset, and the complete training code for full transparency and reproducibility.
V1.2 Improved Version: An updated model with an augmented dataset using LLMs.
Supported Emotions
The model is trained to classify text into 27 distinct emotion categories as well as a neutral class:
You can easily use this model for multi-label emotion classification with the transformers pipeline. Set top_k=None to see all predicted emotions above the model's default threshold.
python
1from transformers import pipeline
23# Load the multilingual, multi-label emotion classification pipeline4emotion_classifier = pipeline(5"text-classification",6 model="multilingual_go_emotions_V1.2",7 top_k=None# To return all scores for each label8)910# --- Example 1: English ---11text_en ="I'm so happy for you, but I'm also a little bit sad to see you go."12results_en = emotion_classifier(text_en)13print(f"Text (EN): {text_en}")14print(f"Predictions: {results_en}\n")1516# --- Example 2: Spanish ---17text_es ="¡Qué sorpresa! No me lo esperaba para nada."18results_es = emotion_classifier(text_es)19print(f"Text (ES): {text_es}")20print(f"Predictions: {results_es}\n")2122# --- Example 3: Arabic ---23text_ar ="أشعر بخيبة أمل وغضب بسبب ما حدث"24results_ar = emotion_classifier(text_ar)25print(f"Text (AR): {text_ar}")26print(f"Predictions: {results_ar}")
Expected Output (structure):
Text (EN): I'm so happy for you, but I'm also a little bit sad to see you go.
Predictions: [[{'label': 'joy', 'score': 0.9...}, {'label': 'sadness', 'score': 0.8...}, {'label': 'caring', 'score': 0.5...}, ...]]
Text (ES): ¡Qué sorpresa! No me lo esperaba para nada.
Predictions: [[{'label': 'surprise', 'score': 0.9...}, {'label': 'excitement', 'score': 0.4...}, ...]]
Text (AR): أشعر بخيبة أمل وغضب بسبب ما حدث
Predictions: [[{'label': 'disappointment', 'score': 0.9...}, {'label': 'anger', 'score': 0.9...}, ...]]
Evaluation
The model's performance was rigorously evaluated on the test set.
Test Set Performance
The following table shows the performance metrics of the fine-tuned model on the test set, broken down by emotion category.
The table below shows the performance of the test model:
Performance of Test Model
Labels
accuracy
precision
recall
f1
mcc
support
threshold
admiration
0.988
0.956
0.942
0.949
0.942
96372
0.60
amusement
0.989
0.951
0.936
0.943
0.937
81726
0.55
anger
0.990
0.938
0.904
0.921
0.915
54456
0.65
annoyance
0.976
0.891
0.856
0.873
0.860
78312
0.50
approval
0.986
0.950
0.903
0.926
0.919
79992
0.65
caring
0.992
0.958
0.936
0.947
0.943
64296
0.55
confusion
0.989
0.941
0.913
0.927
0.921
64638
0.70
curiosity
0.991
0.957
0.937
0.947
0.942
73308
0.70
desire
0.992
0.953
0.934
0.943
0.939
59862
0.55
disappointment
0.979
0.910
0.858
0.883
0.873
75726
0.60
disapproval
0.982
0.900
0.879
0.890
0.880
67158
0.50
disgust
0.992
0.954
0.919
0.936
0.932
54216
0.65
embarrassment
0.994
0.949
0.931
0.940
0.936
44316
0.60
excitement
0.988
0.922
0.899
0.910
0.904
55560
0.60
fear
0.989
0.912
0.904
0.908
0.902
50658
0.50
gratitude
0.996
0.976
0.979
0.978
0.976
74142
0.50
grief
0.992
0.927
0.909
0.918
0.914
39426
0.60
joy
0.983
0.913
0.886
0.899
0.890
72498
0.55
love
0.994
0.967
0.955
0.961
0.957
68226
0.65
nervousness
0.989
0.898
0.872
0.885
0.879
40146
0.60
optimism
0.987
0.949
0.932
0.941
0.933
90498
0.65
pride
0.997
0.955
0.952
0.954
0.952
30918
0.50
realization
0.990
0.956
0.926
0.941
0.935
73908
0.55
relief
0.997
0.966
0.955
0.960
0.959
31728
0.70
remorse
0.995
0.966
0.949
0.957
0.955
49086
0.65
sadness
0.980
0.906
0.872
0.889
0.877
77154
0.55
surprise
0.989
0.928
0.904
0.916
0.910
56130
0.60
neutral
0.981
0.920
0.876
0.898
0.887
79140
0.55
Test Model Performance (Threshold = 0.5)
The table below shows the performance of the test model with a threshold of 0.5:
Labels
accuracy
precision
recall
f1
mcc
support
threshold
admiration
0.988
0.949
0.948
0.949
0.942
96372
0.5
amusement
0.989
0.947
0.938
0.943
0.937
81726
0.5
anger
0.989
0.921
0.917
0.919
0.913
54456
0.5
annoyance
0.976
0.891
0.856
0.873
0.860
78312
0.5
approval
0.986
0.937
0.915
0.926
0.918
79992
0.5
caring
0.992
0.955
0.939
0.947
0.943
64296
0.5
confusion
0.988
0.924
0.927
0.926
0.919
64638
0.5
curiosity
0.990
0.943
0.946
0.945
0.939
73308
0.5
desire
0.992
0.949
0.937
0.943
0.939
59862
0.5
disappointment
0.979
0.894
0.872
0.883
0.871
75726
0.5
disapproval
0.982
0.900
0.879
0.890
0.880
67158
0.5
disgust
0.992
0.942
0.929
0.935
0.931
54216
0.5
embarrassment
0.993
0.941
0.937
0.939
0.935
44316
0.5
excitement
0.988
0.910
0.910
0.910
0.904
55560
0.5
fear
0.989
0.912
0.904
0.908
0.902
50658
0.5
gratitude
0.996
0.976
0.979
0.978
0.976
74142
0.5
grief
0.992
0.916
0.919
0.918
0.913
39426
0.5
joy
0.982
0.907
0.891
0.899
0.889
72498
0.5
love
0.993
0.960
0.960
0.960
0.957
68226
0.5
nervousness
0.989
0.881
0.886
0.884
0.878
40146
0.5
optimism
0.987
0.938
0.941
0.940
0.932
90498
0.5
pride
0.997
0.955
0.952
0.954
0.952
30918
0.5
realization
0.989
0.952
0.928
0.940
0.934
73908
0.5
relief
0.997
0.956
0.963
0.959
0.958
31728
0.5
remorse
0.995
0.957
0.956
0.956
0.953
49086
0.5
sadness
0.979
0.897
0.879
0.888
0.877
77154
0.5
surprise
0.988
0.918
0.912
0.915
0.909
56130
0.5
neutral
0.981
0.914
0.882
0.897
0.887
79140
0.5
Use Cases
This model is ideal for applications requiring nuanced emotional understanding across different languages:
Global Customer Feedback Analysis: Analyze customer reviews, support tickets, and survey responses from around the world to gauge sentiment.
Multilingual Social Media Monitoring: Track brand perception and public mood across different regions and languages.
Advanced Chatbot Development: Build more empathetic and responsive chatbots that can understand user emotions in their native language.
Content Moderation: Automatically flag toxic, aggressive, or sensitive content on international platforms.
Market Research: Gain insights into how different cultures express emotions in text.
Dataset: specialized dataset based on the original Google GoEmotions dataset, augmented using LLMs.
Fine-Tuning Guide
To adapt this model for your own dataset or to replicate the training process, you can follow the methodology outlined in the official code repository. The repository provides a complete, end-to-end example, including data preprocessing, training scripts, and evaluation logic.
For full details, please refer to the GitHub repository:
emotion_chatbot
