patrickjamesmarcellana/synthetic-filipino-sarcasm-detection

Synthetic and Limited Real-World Filipino Sarcasm Detection Dataset

This dataset is composed of Filipino sarcastic and non-sarcastic tweets, divided into two categories: LLM-generated (synthetic) data and real-world data.

Synthetic Data Information

Two large language models were used to generate sarcastic and non-sarcastic tweets for the dataset: GPT-4o and Gemini 2.0 Flash. The dataset is composed of 504 sarcastic tweets (252 for each LLM) and 504 non-sarcastic tweets (252 for each LLM).

Rows

Each row is an instance of a tweet, constrained with X's limitation of 270 characters.

Columns

• —text:
• —the tweet content
• —label:
• —whether the tweet is sarcastic (1) or non-sarcastic (0)
• —possible values: 0, 1
• —sarcasm_type:
• —type of sarcasm according to Encyclopedia Britannica
• —possible values: null, 'self-deprecating', 'brooding', 'deadpan', 'polite', 'obnoxious', 'raging', 'manic'
• —speech_style:
• —type of Filipino speech style used in the tweet
• —possible values: 'general', 'gaylingo', 'conyo', 'balbal'
• —llm_generator:
• —name of LLM used
• —possible values: 'gpt-4o', 'gemini'

Limited Real-World Data Information

Tweets were extracted from the Hate Speech Filipino dataset. Each tweet was manually relabeled to classify each instance according to the presence of sarcasm in the tweets.

Rows

Each row is an instance of a tweet during the 2016 Philippine National Elections

Columns

• —text:
• —the tweet content
• —label:
• —whether the tweet is sarcastic (1) or non-sarcastic (0)
• —possible values: 0, 1