The main goal of this repository is to obtain the phrases or words from a given sentence which strongly agree to the corresponding sentiment. For a detailed information about the tweet sentiment extraction problem and XLNet I strongly suggest this medium post written by me.

Sentiment extraction is natural language processing task where a sentence and it's corresponding sentiment(whether it is a positive or negaitve or neutral sentence) will be given and model is supposed to extract phrases or set of words which strongly support the given sentiment.

The data I used in this repository is from kaggle's tweet sentiment extraction challenge.
In each row ,it has text and corresponding sentiment and selected text.
The challenge is to learn a model to map text ,sentiment to selected text.

We used transformer based models like ALBERT ,XLNET ,ROBERTA to with a CNN head on top.We trained these BERT models along with CNN head using a binary cross entropy model(with 0.1 label smoothing).We tried to predict the start and end indicies ,thereby obtaining the selected text.

We used Stratified K-fold method to reduce the variance in the model.

Jacaard metric is used to evaluate the model.Jacaard metric is quiet similar to IOU in computer vision.To know more about the jaccard metric ,I strongly suggest you to take a look at this awesomely written medium post.

Negative :

Positive :


Neutral :

Negative :

Positive :


Neutral :