Latensis Sentiment — Turkish Sentiment Analysis
Understanding Beyond the Visible
Latensis Sentiment is a Turkish sentiment analysis model fine-tuned on top of Latensis RoBERTa Base — a Turkish-specific RoBERTa model trained from scratch with the Hecemen Unigram 128k tokenizer.
Labels
| ID | Label |
|---|---|
| 0 | Negative |
| 1 | Positive |
Benchmark Results
Evaluated on two independent test sets. Comparison against
savasy/bert-base-turkish-sentiment-cased (BERTurk Sentiment).
| Dataset | Our Model | BERTurk | Notes |
|---|---|---|---|
| TRSAv1 — Accuracy | 0.9027 | 0.8372 | Independent test set |
| TRSAv1 — F1-macro | 0.9026 | 0.8350 | Independent test set |
| Winvoker — Accuracy | 0.8534 | 0.6944 | BERTurk trained on this data |
| Winvoker — F1-macro | 0.7862 | 0.6488 | BERTurk trained on this data |
| SentiTurca Movies — Accuracy | 0.8060 | 0.8186 | Film domain |
On Winvoker test set, our model outperforms BERTurk by +16pp accuracy — notably, BERTurk was trained on Winvoker data while our model had never seen it.
Training Details
| Property | Value |
|---|---|
| Base model | Latensis RoBERTa Base (500k steps, val loss 3.21) |
| Training data | 19k examples (film + product reviews) |
| Labels | Binary (positive / negative) |
| Tokenizer | Hecemen Unigram 128k |
Usage
import torch
import sentencepiece as spm
from transformers import RobertaForSequenceClassification, RobertaConfig
from huggingface_hub import hf_hub_download
# Load tokenizer
spm_path = hf_hub_download(
repo_id="mursideaki/hecemen-tokenizer-unigram-128k",
filename="tr_unigram_tokenizer.model"
)
sp = spm.SentencePieceProcessor()
sp.load(spm_path)
PAD_ID = sp.piece_to_id("<pad>")
BOS_ID = sp.piece_to_id("<s>")
EOS_ID = sp.piece_to_id("</s>")
MAX_LENGTH = 256
def tokenize(text):
ids = sp.encode_as_ids(str(text))
ids = [BOS_ID] + ids[:MAX_LENGTH-2] + [EOS_ID]
mask = [1] * len(ids)
if len(ids) < MAX_LENGTH:
pad_len = MAX_LENGTH - len(ids)
ids = ids + [PAD_ID] * pad_len
mask = mask + [0] * pad_len
return ids, mask
# Load model
config = RobertaConfig.from_pretrained("mursideaki/latensis-sentiment-tr")
model = RobertaForSequenceClassification.from_pretrained(
"mursideaki/latensis-sentiment-tr",
config=config
)
model.eval()
# Predict
def predict(text):
ids, mask = tokenize(text)
input_ids = torch.tensor([ids], dtype=torch.long)
attention_mask = torch.tensor([mask], dtype=torch.long)
with torch.no_grad():
outputs = model(input_ids=input_ids, attention_mask=attention_mask)
label = outputs.logits.argmax(dim=-1).item()
return "positive" if label == 1 else "negative"
print(predict("Bu ürün gerçekten çok kaliteliydi, kesinlikle tavsiye ederim!"))
# → positive
print(predict("Berbat bir deneyimdi, bir daha almam."))
# → negative
Companion Models
- latensis-roberta-base-tr — Base model
- hecemen-tokenizer-unigram-128k — Tokenizer
- NER model (coming soon)
- STS model (coming soon)
Citation
@misc{latensis2026,
author = {Mürşide Aki},
title = {Latensis: Turkish NLP Model Suite},
year = {2026},
publisher = {Hugging Face},
url = {https://huggingface.co/mursideaki/latensis-sentiment-tr}
}
License
MIT