FiliSenti: Unified Filipino Sentiment Model

Model Description

FiliSenti is a fine-tuned XLM-RoBERTa-large (355M parameters) model for ternary sentiment classification (Negative, Neutral, Positive) of Filipino text, combining Tagalog and Hiligaynon in a single model. It was trained on TagaSenti (Tagalog, 35,686 sentences) and HiliSenti (Hiligaynon, 21,095 sentences), for 56,781 sentences in total.

The model achieves 89.2% test accuracy and a macro F1 of 0.8915 on the combined held-out test set, the highest published score for 3-class Filipino sentiment analysis. Per-language performance is 91.7% F1 on Hiligaynon and 88.1% F1 on Tagalog, meaning the same model serves both languages without any language switch.

A quantized INT8 ONNX version of the model is included for on-device deployment (Flutter + ONNX Runtime) with no internet connection required after the model is placed.

Intended Uses

• Sentiment analysis of Filipino social-media and customer feedback
• Cross-lingual / unified analysis across Tagalog- and Hiligaynon-speaking communities (Negros Occidental, Panay Island, Soccsksargen, and the wider Philippines)
• On-device, privacy-preserving sentiment analysis on Android devices
• Research on low-resource Philippine languages and cross-lingual transfer

Out-of-Scope Uses

• Fine-grained emotion detection
• Aspect-based sentiment analysis
• High-stakes decision-making without human verification
• Languages other than Tagalog and Hiligaynon (the model may generalize to some extent via XLM-RoBERTa's multilingual base, but is not optimized for them)

Training Data

The model was trained on a combination of TagaSenti (Tagalog) and HiliSenti (Hiligaynon):

Combined stratified split (80 / 10 / 10, by label). Note that the HiliSenti contribution is its train and validation splits only; its test split is excluded entirely, so the totals below do not include it:

The test split was fully held out during model development and never used for hyper-parameter tuning.

Training Procedure

• Base model: xlm-roberta-large (355M parameters)
• Tokenizer: XLM-RoBERTa-large SentencePiece tokenizer with vocabulary augmentation, in which frequently fragmented Filipino words were added as new tokens (final vocab size 250,011)
• Text normalisation: NFKC, laughter collapsing, reduplication (ganda-ganda → ganda ganda, sobra2 → sobra sobra), slang expansion for Tagalog (wla→wala, dko→di ko, kc→kasi) and Hiligaynon (gd→gid, wla→wala, sbng→subong), repeated-character collapse. Casing preserved, since XLM-RoBERTa is a cased model
• Max sequence length: 128 tokens
• Batch size: 16 (effective batch size 32 via gradient accumulation)
• Learning rate: 2e-5 with cosine schedule (min LR 1e-6) and 312 warm-up steps
• Optimizer: AdamW (fused)
• Epochs: 5 (early stopping with patience 3; best val F1 0.8895 @ step 5400)
• Label smoothing: 0.1
• Class weighting: Inverse-frequency weights applied to cross-entropy loss
• Mixed precision: FP16 · Gradient checkpointing: enabled
• Compute: Google Colab (Tesla T4 GPU)
• Export: FP32 → ONNX (opset 18, dynamic batch/sequence), then INT8 dynamic quantization (per_channel=True) in an isolated subprocess to control memory, producing a single-file filisenti_int8.onnx (~537 MB)

Evaluation Results

Combined test set (5,678 sentences)

Per-language (test set)

These per-language figures are not directly comparable with the 93.4% macro F1 reported by the HiliSenti-only model. FiliSenti re-partitions the combined corpus into its own splits, so its Hiligaynon test portion (2,109 sentences) is drawn from the HiliSenti train and validation splits, which the HiliSenti-only model was itself trained on. The 91.7% figure is therefore not a measurement of accuracy retained on unseen Hiligaynon data. The trade-off offered here is coverage rather than accuracy: one checkpoint serving both languages (91.7% Hiligaynon, 88.1% Tagalog) instead of Hiligaynon alone.

Limitations

• Dialectal bias: Hiligaynon training data is weighted toward Negros Occidental; Panay and Soccsksargen varieties are under-represented. Tagalog data skews toward the e-commerce, news, and social-media domains of TagaSenti.
• Task scope: Only sentence-level ternary sentiment is supported.
• Sarcasm: Despite targeted adversarial augmentation in TagaSenti, sarcasm detection remains imperfect.
• LLM-annotated data: Parts of TagaSenti were labeled by an LLM without validation against human ground truth; annotation biases may propagate.
• Two-language coverage: The model covers Tagalog and Hiligaynon only; other Philippine languages are not benchmarked.

Ethical Considerations

• Training data comes from publicly available sources (news, public social media) and the openly published TagaSenti / HiliSenti datasets.
• The source datasets contain real names, locations, and descriptions of events, and have not been anonymized. Apply your own anonymization or filtering pipeline where required by institutional or regulatory privacy rules.
• Non-commercial restrictions on the underlying datasets apply to derivative models (see Licenses).

How to Use

Via Transformers

from transformers import AutoModelForSequenceClassification, AutoTokenizer

model = AutoModelForSequenceClassification.from_pretrained("jjjardev/filisenti")
tokenizer = AutoTokenizer.from_pretrained("jjjardev/filisenti")

sentence = "Sobrang sarap ng pagkain dito sa restaurant na ito."
inputs = tokenizer(sentence, return_tensors="pt")
outputs = model(**inputs)
prediction = outputs.logits.argmax().item()

labels = ["Negative", "Neutral", "Positive"]
print(labels[prediction])  # Positive

Using the Pipeline API

from transformers import pipeline

classifier = pipeline("text-classification", model="jjjardev/filisenti")
result = classifier("Wala na talaga, ayaw ko na dito. Ang sama ng ugali.")
print(result)  # [{'label': 'Negative', 'score': ...}]

On-Device (ONNX)

Download the INT8 ONNX model: https://huggingface.co/jjjardev/filisenti/resolve/main/filisenti_int8.onnx

import onnxruntime as ort
from transformers import AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("jjjardev/filisenti")
session = ort.InferenceSession("filisenti_int8.onnx", providers=["CPUExecutionProvider"])

text = "Salamat gid sa inyo bulig!"
inputs = tokenizer(text, return_tensors="np", truncation=True, max_length=128)
logits = session.run(["logits"], {
    "input_ids": inputs["input_ids"],
    "attention_mask": inputs["attention_mask"],
})[0]
print(labels[logits.argmax().item()])

ONNX spec: inputs input_ids (int64) and attention_mask (int64) with dynamic [batch, seq_len]; output logits (float32 [batch, 3]). This file is what the FiliSenti Flutter app and web demo load.

Citation

If you use this model or the TagaSenti / HiliSenti datasets in your research, please cite:

@misc{jarder2026filisenti,
  author    = {Jarder, Jessie James T.},
  title     = {FiliSenti: A Unified 3-Class Filipino Sentiment Model for Tagalog and Hiligaynon},
  year      = {2026},
  publisher = {Hugging Face},
  doi       = {10.57967/hf/9796},
  url       = {https://huggingface.co/jjjardev/filisenti}
}

Dataset references:

@dataset{jarder2026tagasenti,
  author    = {Jessie James T. Jarder},
  title     = {TagaSenti: A Multi-Domain Sentiment Analysis Dataset for Tagalog},
  year      = {2026},
  publisher = {Hugging Face},
  doi       = {10.57967/hf/9620},
  url       = {https://huggingface.co/datasets/jjjardev/tagasenti}
}

@dataset{jarder2026hilisenti,
  author    = {Jessie James T. Jarder},
  title     = {HiliSenti: A Multi-Domain Sentiment Analysis Dataset for Hiligaynon},
  year      = {2026},
  publisher = {Hugging Face},
  doi       = {10.57967/hf/8737},
  url       = {https://huggingface.co/datasets/jjjardev/hilisenti-v1}
}

Licenses

Related Resources

• Datasets: jjjardev/tagasenti · jjjardev/hilisenti-v1
• Code: github.com/jjjardev/filisenti
• ONNX model: filisenti_int8.onnx in this repository

Contact

Jessie James T. Jarder, jj.jarder.dev@gmail.com