paraphrase-albert-small-v2 fine-tuned for emotion classification

A small ALBERT-based Classifier that assigns one of six Emotions to a short English text: sadness, joy, love, anger, fear or surprise. It is sentence-transformers/paraphrase-albert-small-v2 fine-tuned on dair-ai/emotion.

Code, training notebook and API: https://github.com/PierreBoissier/emotion_classification

Usage

from transformers import pipeline

classifier = pipeline("text-classification", model="pboissie/paraphrase-albert-small-v2-emotion", top_k=None)
classifier("i am so happy to see you again today")

Texts longer than 256 tokens are truncated.

Results

Evaluated on the 2,000-text test Split of dair-ai/emotion:

Metric	Value
Accuracy	0.921
Macro F1	0.875

Emotion	Precision	Recall	F1	Support
sadness	0.960	0.952	0.956	581
joy	0.937	0.948	0.943	695
love	0.813	0.792	0.803	159
anger	0.911	0.931	0.921	275
fear	0.869	0.920	0.894	224
surprise	0.875	0.636	0.737	66

The weakest Emotion is surprise (recall 0.636), mostly mistaken for fear. Love is often mistaken for joy.

Training

Data: the dair-ai/emotion train Split (16,000 texts) for training; the validation Split (2,000 texts) is only used to monitor accuracy, not to train or select weights.
10 epochs, learning rate 2e-5, weight decay 0.01, batch size 32 with 4 gradient accumulation steps, AdamW with a linear schedule (5% warmup), mixed precision, on a Colab GPU.
The weights of the last epoch are kept. Validation accuracy reaches about 0.93 and is flat over the last epochs.

Limitations

English only, and trained on short texts: the dataset is made of English tweets labelled from hashtags, so labels are noisy and the Emotions reflect that style of writing.
Classes are imbalanced (surprise and love are rare), which explains their lower scores.
Each text gets exactly one Emotion; mixed or neutral feelings are not modelled.
Not intended for decisions about people: it is a portfolio project, not a validated tool.