Text Classification Competition: Twitter Sarcasm Detection

Repository Description

• /BERT-base-uncase-code:
  Everything you need to test the BERT model
• /CNN_LSTM_DNN:
  Everything you need to test the neural-network-based model
• /ML:
  Everything you need to test the machine learning model
• /data:
  Raw Data
• /resource:
  Supportive documents for neural-network-based model
• /CS 410 Text Information Systems Course Project Final Report.docx:
  Final Report
• /answer.txt：
  Best output we have
• /slides.pptx：
  Slides, a brief version of Report

Important Notes (Read before testing the model):

We highly recommend you to contact us for a live demo before you test the model on your own. Some pre-requisites are hard to meet which may lead to errors during the test. The testing process is also highly time-consuming. In order to better display our results and to save your time, please contact us via the following email address:

hh21@illinois.edu

Or you can this live demo.

Thank you for your understanding!

How to test models on your PC?

We performed three different models for this competition, machine learning model, BERT-based model and neural-network model. You can test whichever you like by following the instructions respectively.

Machine Learning Model

Pre-requisite:

• numpy
• pandas
• matplotlib
• seaborn
• sklearn

Run the script:

1. Clone the repository to your computer
2. cd ClassificationCompetition/ML
3. Run python TFIDF_RandomForests.py
4. Output prediction result (anser.txt) will be able to find in ClassificationCompetition folder

BERT-base-uncased Model

Pre-requisite:

• Tensorflow
• Transformer
• PyTorch
• torchtext
• BERT
• cuda toolkit https://anaconda.org/anaconda/cudatoolkit (Hardware requirments: https://docs.anaconda.com/anaconda/user-guide/tasks/gpu-packages/)
• numpy
• pandas
• matplotlib
• seaborn
• sklearn

Run the script:

1. Clone the BERT-base-uncase-code into your local computer (all the dataset is already prepared for this model)
2. Run python BERT_Model.py (If you see this message in the console "Running this sequence through the model will result in indexing errors", this is just the warning, NOT actual error!)
3. The final models (metrics.pt and model.pt) and output prediction result (anser.txt) will be able to find in result folder

CNN+LSTM+DNN Model

Pre-requisite:

• nltk (TweetTokenizer)
• Keras
• Tensorflow
• numpy
• scipy
• gensim (if you are using word2vec)
• itertools
• sklearn

Run the script:

1. Clone the repository
2. Download following files from the link - https://drive.google.com/drive/folders/0B7C_0ZfEBcpRbDZKelBZTFFsV0E?usp=sharing, to the following directory: ClassificationCompetition/resource/text_model/weights
   • GoogleNews-vectors-negative300.bin
   • model.jsonhdf5
   • weights.05__.hdf5
3. cd ClassificationCompetition/CNN_LSTM_DNN
4. run Python sarcasm_detection_model_CNN_LSTM_DNN.py