Sarcasm Detection Model

A bidirectional LSTM that classifies news headlines as sarcastic or sincere, built for a hackathon. Sarcasm is hard precisely because the words say one thing and mean another.

Most sarcasm datasets come from social media, labelled by hashtags — noisy labels, and many posts are replies that need their conversation to make sense. A dataset of satirical and genuine news headlines avoids both: each headline stands alone, and the label comes from where it was published.

1. Clean

   Lowercase, expand contractions, strip punctuation, lemmatise.

2. Encode

   Tokenise, pad to the longest headline, map to pretrained GloVe vectors, kept frozen.

3. Model

   A bidirectional LSTM reads each headline both ways; global max pooling keeps the strongest signal.

4. Classify

   Two dense layers with heavy dropout, then a single sigmoid output.

5. Train

   Early stopping and learning-rate reduction on a validation split.

Split

80/20 train/test, with validation carved from training.

Baseline

Classes are near-balanced, so always guessing the majority scores roughly half.

Balance

Precision and recall land within a point of each other on both classes — the accuracy is not bought by favouring one.

The label is which site published the headline, so the model can learn a publication's style as well as sarcasm itself — it has never been tested on headlines from anywhere else. Results come from a single random split, not cross-validation, and training accuracy runs a few points above validation, a mild overfit. On review I also found the 50-dimension embeddings were tiled four times to fill a 200-dimension layer, which adds size without adding information.

• Use embeddings at their native size instead of tiling them. First fix, and a one-line change.
• Test on headlines from other publications, to see whether it learned sarcasm or a house style.
• Cross-validate, and compare against a fine-tuned transformer as the obvious stronger baseline.