Fig. 4

Different examples of correctly classified sequences from the IMDB dataset. From left to right: (a): Typical example of a positively classified movie (blue). The first word that clearly classifies the sequence as positive is ‘powerful’. Afterward, no word suggests that the movie might have been negatively rated. The words ‘wonderfully’, ‘strong’, and ‘fascinating’ result in particularly larger jumps in the hidden state space; (b): Typical example of a negatively classified sequence (pink). Right from the beginning, this review was classified negatively; (c): The input starts with several negative words (pink), making the model alternate between hidden states with high and low EPs. Toward the end of the review, the intent of the writer becomes clear and the model settles for a positive output (blue); (d): A positively classified movie with uncertainties. At the beginning and end, there are more indications of a positive rating (blue). However, most vocabulary feels more neutral than highly positive. In the middle, there is also a negative statement (pink). Underlying data source: IMDB as available in Keras [13]