🤖 Sentiment Classifier with Sarcasm Detection
This model combines roberta-base embeddings with a BiLSTM layer and attention mechanism to classify the sentiment of movie reviews, even when the language includes sarcasm, irony, or mixed emotional signals.
🧩 Architecture
- Base Encoder: RoBERTa (frozen)
- Sequence Modeling: Bidirectional LSTM (BiLSTM)
- Attention Layer: Learnable scalar attention weights
- Classifier: Fully connected + Softmax
This model takes the output of RoBERTa, processes it through a BiLSTM, applies attention to focus on important parts of the sentence, and classifies sentiment as Positive (1) or Negative (0).
💡 Why this model?
Standard BERT-based sentiment classifiers often fail on subtle language such as:
- "If I had a dollar for every cliché in this film, I’d be rich." → sarcastic → Negative
- "It’s fine. Just fine. Totally… fine." → sarcastic tone → Negative
- "The visuals were stunning, but the plot was as dead as a doorknob." → contrast → Negative
This model aims to address that by combining contextual language modeling (RoBERTa) with sequence learning (BiLSTM) and attention-based interpretability.
📊 Training Details
- Dataset: IMDB (2-class, 25k reviews)
- Training Subset: 2,000 samples for prototype
- Loss: CrossEntropyLoss
- Optimizer: Adam
- Epochs: 3
- Hardware: Google Colab (T4 GPU)
- Additional Inputs: Custom sarcastic and ambiguous examples
🧪 Example Inputs
| Input | Prediction |
|---|---|
| "If I had a dollar for every cliché..." | 🔴 Negative |
| "It’s fine. Just fine. Totally fine." | 🔴 Negative |
| "The acting was great, but the plot..." | 🔴 Negative |
| "Brilliant visuals and solid pacing." | 🟢 Positive |
🚀 How to Use
from transformers import RobertaTokenizer
from my_model import RobertaBiLSTMAttention # the custom model
import torch
tokenizer = RobertaTokenizer.from_pretrained("roberta-base")
model = RobertaBiLSTMAttention()
model.load_state_dict(torch.load("pytorch_model.bin"))
model.eval()
text = "It’s fine. Just fine. Totally… fine."
tokens = tokenizer(text, return_tensors="pt", padding="max_length", truncation=True, max_length=128)
logits = model(tokens["input_ids"], tokens["attention_mask"])
pred = torch.argmax(logits, dim=1).item()
print("Sentiment:", "Positive" if pred == 1 else "Negative")