IronySarcasmDetectorIT

NLP project that analyses Italian tweets and finds out if they are ironic or not, and if they're ironic, if they are sarcastic or not.

Challenge

The tasks solved in this project are the ones from the EVALITA 2018 challenge. The detailed description can be found here:

@inproceedings{cignarella2018overview,
  title={Overview of the {EVALITA} 2018 task on Irony Detection in {I}talian tweets ({IronITA})},
  author={Cignarella, Alessandra Teresa and Frenda, Simona and Basile, Valerio and Bosco, Cristina and Patti, Viviana and Rosso, Paolo and others},
  booktitle={Sixth Evaluation Campaign of Natural Language Processing and Speech Tools for Italian (EVALITA 2018)},
  volume={2263},
  pages={1--6},
  year={2018},
  organization={CEUR-WS}
}

Model

The framework used to develop the models is tensorflow.

Among all the models we developed, the one producing the best results has been the BERT-GRU model.

The BERT model we used was a pretrained one available on HuggingFace (https://huggingface.co/dbmdz/bert-base-italian-xxl-cased).

Trained BERT-GRU model

Here you can download the trained BERT-GRU model (≈ 500MB). Another 400MB will be downloaded by the get_bert_tokenizer which will downloaded a pretrained tokenizer from HuggingFace.co

Here is an example of how to use it:

from models.bert_tokenizer import get_bert_tokenizer, tokenize
from preprocessing.pipeline import ItalianTweetsPreprocessingPipeline

pp = ItalianTweetsPreprocessingPipeline()

model = keras.models.load_model('path/to/trained/model/location')

tokenizer = get_bert_tokenizer()

tweets = pp.apply(pandas_frame_containing_tweets)
tokenized_tweets = tokenize(tweets, tokenizer)[:-1]

pred = model.predict(tokenized_tweet)

Challenge Results

• Task A

• Task B

Code structure

The main code is split across there three folders:

• preprocessing - contains utilities for text preprocessing implementing the pipeline through which the tweets will pass.
• models - contains the models implementations.
• notebooks - contains the Jupyter notebooks we used to develop and test the code. The notebooks containing the models' implementations are meant to be run on Colab.