You build a sentiment analysis system that feeds word tokens…
You build a sentiment analysis system that feeds word tokens into a unidirectional RNN, then outputs the sentiment class label by sending the final hidden state through a linear + SoftMax layer. You observe that your model incorrectly predicts very positive sentiment for the following (negative sentiment) passage: “The play was terrible. The performances were lackluster and the acting was unconvincing. Then there was a long line to exit the theatre building. At least the venue was nice and the dinner was excellent.” Why might the model make this decision?