MarcelM/CSEBERT-Emotion-Detection

CSEBERT for Ekman emotion recognition in Slovene

Multi-label emotion classifier for Slovene text, covering Ekman's six basic emotions plus a neutral label.

The model is CroSloEngual BERT fine-tuned on the English GoEmotions corpus with labels mapped to Ekman's taxonomy, then applied to Slovene. No Slovene training data was used — the model relies on CroSloEngual BERT's shared Slovene/Croatian/English representation to transfer the task across languages.

Produced as part of an undergraduate thesis at the Faculty of Computer and Information Science, University of Ljubljana.

Labels

Seven independent binary outputs, not a distribution. A comment may carry several labels or none.

id	label	Slovene
0	anger	jeza
1	disgust	gnus
2	fear	strah
3	joy	sreča
4	sadness	žalost
5	surprise	presenečenje
6	neutral	nevtralno

The GoEmotions 27-label taxonomy was collapsed to Ekman using the official ekman_mapping.json grouping.

Usage

import torch
from transformers import AutoModelForSequenceClassification, AutoTokenizer

MODEL_ID = "<uporabnik>/csebert-emotions-sl"
tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
model = AutoModelForSequenceClassification.from_pretrained(MODEL_ID).eval()

text = "Kako ogabno pripravljena hrana... kje ti je higiena??"
with torch.no_grad():
    enc = tokenizer(text, truncation=True, max_length=128, return_tensors="pt")
    probs = torch.sigmoid(model(**enc).logits)[0]

for i, p in enumerate(probs):
    print(f"{model.config.id2label[i]:12s} {p:.3f}")

Reported results use a decision threshold of 0.5.

Training

Base model	EMBEDDIA/crosloengual-bert
Training data	GoEmotions, English split, Ekman mapping
Objective	multi-label, BCEWithLogitsLoss
Learning rate	3e-05
Batch size	16
Epochs	10, early stopping with patience 3 on validation macro-F1
Max sequence length	128
Precision	fp16
Seed	42

Selected from a grid over learning rate {1e-05, 2e-05, 3e-05, 5e-05} and batch size {16, 32, 64} by validation macro-F1.

Input preprocessing

This matters for getting comparable results. The evaluation data was preprocessed before inference:

• named entities replaced with [NAME]
• digit strings replaced with [NUMBER], currency amounts with [CURRENCY]
• URLs, <3 and leading > removed, & expanded to in
• ASCII art removed, comments kept to 3–30 tokens

Raw user text will not match this distribution. Expect somewhat lower quality on input that has not been through the same cleanup, particularly text containing names or numbers.

Intended use

Research and demonstration: exploratory analysis of emotion in Slovene social media text, and as a baseline for further work on Slovene emotion recognition.

Not intended for decisions about individuals — moderation, hiring, clinical or psychological assessment, or anything where a wrong label affects someone.

Attribution

• Base model: CroSloEngual BERT, Ulčar and Robnik-Šikonja
• Training data: GoEmotions, Demszky et al. (2020)