T5 Base Finetuned Sarcasm Twitter

🚀 T5-base fine-tuned for Sarcasm Detection 🙄

A fine-tuned T5-base model on the Twitter Sarcasm Dataset for sequence classification as text generation.

🚀 Quick Start

This is a T5-base model fine-tuned on the Twitter Sarcasm Dataset for the Sequence classification (as text generation) downstream task.

from transformers import AutoTokenizer, AutoModelWithLMHead
tokenizer = AutoTokenizer.from_pretrained("mrm8488/t5-base-finetuned-sarcasm-twitter")

model = AutoModelWithLMHead.from_pretrained("mrm8488/t5-base-finetuned-sarcasm-twitter")

def eval_conversation(text):

  input_ids = tokenizer.encode(text + '</s>', return_tensors='pt')

  output = model.generate(input_ids=input_ids, max_length=3)
  
  dec = [tokenizer.decode(ids) for ids in output]

  label = dec[0]

  return label

# For similarity with the training dataset we should replace users mentions in twits for @USER token and urls for URL token.

twit1 = "Trump just suspended the visa program that allowed me to move to the US to start @USER!" +
" Unfortunately, I won’t be able to vote in a few months but if you can, please vote him out, " +
"he's destroying what made America great in so many different ways!"

twit2 = "@USER @USER @USER We have far more cases than any other country, " +
"so leaving remote workers in would be disastrous. Makes Trump sense."

twit3 = "My worry is that i wouldn’t be surprised if half the country actually agrees with this move..."

me = "Trump doing so??? It must be a mistake... XDDD"

conversation = twit1 + twit2

eval_conversation(conversation) #Output: 'derison'

conversation = twit1 + twit3

eval_conversation(conversation) #Output: 'normal'

conversation = twit1 + me

eval_conversation(conversation) #Output: 'derison'

# We will get 'normal' when sarcasm is not detected and 'derison' when detected

✨ Features

• Fine-tuned on the Twitter Sarcasm Dataset for sarcasm detection.
• Converts sequence classification into a text generation task.

📚 Documentation

Details of T5

The T5 model was presented in Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer by Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, Peter J. Liu.

Transfer learning, where a model is first pre-trained on a data-rich task before being fine-tuned on a downstream task, has emerged as a powerful technique in natural language processing (NLP). The effectiveness of transfer learning has given rise to a diversity of approaches, methodology, and practice. In this paper, the authors explore the landscape of transfer learning techniques for NLP by introducing a unified framework that converts every language problem into a text-to-text format. Their systematic study compares pre-training objectives, architectures, unlabeled datasets, transfer approaches, and other factors on dozens of language understanding tasks. By combining the insights from their exploration with scale and their new “Colossal Clean Crawled Corpus”, they achieve state-of-the-art results on many benchmarks covering summarization, question answering, text classification, and more. To facilitate future work on transfer learning for NLP, they release their dataset, pre-trained models, and code.

Details of the downstream task (Sequence Classification as Text generation) - Dataset 📚

The Twitter Sarcasm Dataset provides training and testing datasets for sarcasm detection tasks in jsonlines format.

Each line contains a JSON object with the following fields:

• label: SARCASM or NOT_SARCASM
  • NOT in test data
• id: String identifier for sample. This id will be required when making submissions.
  • ONLY in test data
• response: the sarcastic response, whether a sarcastic Tweet
• context: the conversation context of the response
  • Note, the context is an ordered list of dialogue. If the context contains three elements, c1, c2, c3, in that order, then c2 is a reply to c1 and c3 is a reply to c2. Further, if the sarcastic response is r, then r is a reply to c3.

For example, for the following training example:

"label": "SARCASM", "response": "Did Kelly just call someone else messy? Baaaahaaahahahaha", "context": ["X is looking a First Lady should . #classact", "didn't think it was tailored enough it looked messy"]

The response tweet, "Did Kelly..." is a reply to its immediate context "didn't think it was tailored..." which is a reply to "X is looking...". The goal is to predict the label of the "response" while also using the context (i.e., the immediate or the full context).

Dataset size statistics:

The datasets were preprocessed to convert them to a text-to-text (classification as a generation task).

Model fine-tuning 🏋️‍

The training script is a slightly modified version of this Colab Notebook created by Suraj Patil. All credits go to him!

Test set metrics 🧾

🔧 Technical Details

The model is based on Google's T5-base architecture and fine-tuned on the Twitter Sarcasm Dataset. The fine-tuning process involves converting the sequence classification task into a text generation task, which allows the model to leverage the power of the T5's text-to-text framework.

📄 License

No license information is provided in the original README.

Created by Manuel Romero/@mrm8488 | LinkedIn

Made with ♥ in Spain