All the code created for our participation in the Rest-Mex 2022 Conference is available at this GitHub Repository.
- Gabriel Missael Barco.
- Gil Estéfano Rodríguez Rivera.
- Delia Irazú Hernández Farías.
We participated in the sentiment analysis track, classifying tourism reviews into two domains: the polarity measured by the number of stars assigned in the review (from 1 to 5) and the type of attraction being reviewed (namely Hotel, Restaurant, or Attractive). For doing this, the conference organizers provided us with a labeled data-set. Given that the data-set was imbalanced, we extracted data from the Yelp Open Dataset to balance the data-set and performed translation from English to Spanish.
We used a pre-trained BERT model for the polarity sub-task (nlptown/bert-base-multilingual-uncased-sentiment), and we fine tuned the model on the data. Our model excelled in classifying the low stars reviews (1 and 2) because of the data augmentation for balancing. We used a bag-of-words model with a classification head using different machine learning algorithms for the attraction type sub-task. We obtained good results in this sub-task while keeping the computational cost lower than if we were using a BERT-based model for this sub-task.
Disclaimer. The notebooks available in this repository are not meant to be executed; they are meant to be read-only, as a reference. The train data and the test data are not available in this repository.
In progress...
Data augmentation:
Model evaluation for polarity sub-task:
Second approach for the polarity sub-task:
