Interactive Visualizations for Deep Learning

Fig. 1. This SentimentTree visualization shows how word-level sentiments build up over multi-word phrases to produce the overall sentiment associated with a sentence. Leaf nodes represent individual words such as Steven, Solaris, and failure. Non-terminal nodes represent phrases such as glorious failure. The root node represents the full sentence: If Steven Soderbergh’s ‘Solaris’ is a failure it is a glorious failure. Colors encode sentiment ratings given by human subjects, where blue is positive and red is negative. Human raters find the majority of words in this sentence to be neutral with the exceptions of failure being negative and glorious being positive. Sentiment compositionality comes from combining words or phrases to form longer text, but the resulting ratings can take many forms. The phrase glorious failure is rated as negatively as the word failure itself. However, while the two main clauses of this sentence, Steven Soderbergh’s ‘Solaris’ is a failure and it is a glorious failure are both negatively rated, their concatenation suggests an overall positive view about the film Solaris. Our algorithm produces the current state-of-the-art accuracy on sentence-level sentiment predictions.