Sengil/Turkish-ABSA-BiLSTM-Word2Vec

🇹🇷 Turkish Aspect-Based Sentiment Analysis (ABSA) – BiLSTM + Word2Vec

This model performs aspect-based sentiment analysis (ABSA) on Turkish sentences. Given a sentence and a specific aspect, it predicts the sentiment polarity (Negative, Neutral, Positive) associated with that aspect.

🧠 Model Details

• Model Type: BiLSTM (Bidirectional Long Short-Term Memory) + Word2Vec
• Developer: Sengil
• Library: Keras
• Input Format: "Sentence [ASP] Aspect"
• Labels: 0 = Negative, 1 = Neutral, 2 = Positive
• Training Dataset: Sengil/Turkish-ABSA-Wsynthetic

📊 Evaluation Results

The model achieved the following performance on the test set:

Class	Precision	Recall	F1-Score	Support
Negative	0.89	0.91	0.90	896
Neutral	0.70	0.64	0.67	140
Positive	0.92	0.92	0.92	1178
Overall			0.90	2214

• Overall Accuracy: 90%
• Macro-Averaged F1-Score: 83%
• Weighted-Averaged F1-Score: 90%

🚀 Usage Example

Download model from HF

from huggingface_hub import hf_hub_download
import pickle
from tensorflow.keras.models import load_model

model_path = hf_hub_download(repo_id="Sengil/Turkish-ABSA-BiLSTM-Word2Vec", filename="absa_bilstm_model.keras")
tokenizer_path = hf_hub_download(repo_id="Sengil/Turkish-ABSA-BiLSTM-Word2Vec", filename="tokenizer.pkl")

# load model
model = load_model(model_path)

# load tokenizer
with open(tokenizer_path, "rb") as f:
    tokenizer = pickle.load(f)

Input preprocessing

import re
import nltk
nltk.download('punkt')

def preprocess_turkish(text):
    text = text.lower()
    text = re.sub(r"http\S+|www\S+|https\S+", "<url>", text)
    text = re.sub(r"@\w+", "<user>", text)
    text = re.sub(r"[^a-zA-Z0-9çğıöşüÇĞİÖŞÜ\s]", " ", text)
    text = re.sub(r"(.)\1{2,}", r"\1\1", text)
    text = re.sub(r"\s+", " ", text).strip()
    return text

Predict the input

import numpy as np
from tensorflow.keras.preprocessing.sequence import pad_sequences

def predict_sentiment(sentence, aspect, max_len=84):
    input_text = sentence + " [ASP] " + aspect
    cleaned = preprocess_turkish(input_text)
    tokenized = tokenizer.texts_to_sequences([cleaned])
    padded = pad_sequences(tokenized, maxlen=max_len, padding='post')
    
    pred = model.predict(padded)
    label = np.argmax(pred)
    labels = {0: "Negatif", 1: "Nötr", 2: "Pozitif"}
    return labels[label]

run

sentence = "Manzara sahane evet ama servis rezalet."
aspect = "manzara"

predict = predict_sentiment(sentence, aspect)
print("predict:", predict)

🏋️‍♀️ Training Details

• Embedding: Word2Vec (dimension: 100)

• Model Architecture:

  • Embedding layer (initialized with pre-trained Word2Vec weights)
  • 2 x BiLSTM layers (each with 100 units, dropout: 0.3)
  • Conv1D layer (100 filters, kernel size: 5)
  • Global Max Pooling
  • Dense layer (16 units, ReLU activation)
  • Output layer (3 units, softmax activation)

• Training Parameters:

  • Loss Function: sparse_categorical_crossentropy
  • Optimizer: Adam
  • Epochs: 35 (with early stopping)
  • Batch Size: 128
  • Learning Rate: 1e-3 (adjusted dynamically with ReduceLROnPlateau)

📚 Training Data

The model was trained on the Sengil/Turkish-ABSA-Wsynthetic dataset, which comprises semi-synthetic Turkish sentences annotated for aspect-based sentiment analysis, particularly in the restaurant domain.

⚠️ Limitations

• Performance on the Neutral class is lower compared to other classes, possibly due to class imbalance in the training data.
• The model may struggle with rare or ambiguous aspects not well represented in the training set.
• Complex sentence structures or ironic expressions may affect the model's accuracy.

📄 Citation

@misc{turkish_absa_bilstm_word2vec,
  title  = {Turkish Aspect-Based Sentiment Analysis using BiLSTM + Word2Vec},
  author = {Sengil},
  year   = {2025},
  url    = {https://huggingface.co/Sengil/Turkish-ABSA-BiLSTM-Word2Vec}
}

📬 Contact

For questions or feedback, please reach out via Hugging Face profile.