[README UNDER CONSTRUCTION]
emBert is a Hungarian text classification model, aimed at classifying 7 possible emotions and a neutral state. The model uses
huBERT
tokenizer, and was fine-tuned on a
huBERT
base model with a proprietary database of Hungarian online news site sentences. The sentences for the fine-tuning set were classified manually by experts in a double-blind manner. Inconsistencies were dealt with manually.
The results of the fine-tuning validation were:
|
emotion
|
precision
|
recall
|
f1-score
|
|
0 - Anger
|
0.70
|
0.74
|
0.72
|
|
1 - Disgust
|
0.72
|
0.73
|
0.73
|
|
2 - Fear
|
0.61
|
0.47
|
0.53
|
|
3 - Happiness
|
0.38
|
0.37
|
0.38
|
|
4 - Neutral
|
0.65
|
0.62
|
0.63
|
|
5 - Sad
|
0.74
|
0.72
|
0.73
|
|
6 - Successful
|
0.79
|
0.81
|
0.80
|
|
7 - Trustful
|
0.76
|
0.78
|
0.77
|
|
weighted avg
|
0.73
|
0.74
|
0.73
|
|
Accuracy reached 74%.
|
|
|
|
The emotions are based on
Plutchik 1980
, with anticipation substituted with neutral.
Proper use of the model:
from transformers import AutoTokenizer, AutoModelForSequenceClassification
tokenizer = AutoTokenizer.from_pretrained("SZTAKI-HLT/hubert-base-cc")
model = AutoModelForSequenceClassification.from_pretrained("poltextlab/emBERT")
The model was created by György Márk Kis, Orsolya Ring, Miklós Sebők of the Center for Social Sciences.