This repo contains the data, scripts, and results from the SemEval 2019/2020 Sentimix Spanglish challenge.
The challenge deals with the problem of identifying the sentiment of a set of 'Spanglish' tweets, where both English and Spanish are used in a single tweet. Misspellings, new words, mixed grammar, and the short length of tweets make this task difficult, and currently prevailing methods of using pre-trained contextual word embeddings may not be as effective. Vocabulary and embedding sizes will have to be large to accommodate both languages, and the odds of running into an out-of-vocabulary word are very high.
We have ran experiments using one-hot encoded character embeddings and concatenated Spanish + English word embeddings.
Spanish + English Word Embeddings 3 CNN layers, Max Pooling, 2 Dense Layers for classification
| Precision | Recall | F1 |
|---|---|---|
| 0.61 | 0.3348 | 0.4323 |
One-Hot Encoded Character Embeddings 1 CNN layer, Max Pooling, 2 Dense Layers for classification
| Precision | Recall | F1 |
|---|---|---|
| 0.42 | 0.56 | 0.41 |
Corey Sutphin - Preprocessing scripts, model utilizing Spanish word embeddings, English word embeddings, and then the two concatenated on each other. Cove Soyars - Preprocessing scripts, bash script, model using one-hot encoded character embeddings with a CNN. Nick Rodriguez - Preprocessing scripts, BiLSTM model