Visualizing and Understanding Neural Models in NLP Jiwei Li, Xinlei Chen, Eduard Hovy and Dan Jurafsky Presentation by Rohit Gupta Roll No. 15111037
Motivation Vector-based models produced by applying neural networks to natural language are very difficult to interpret
Dataset used Stanford Sentiment Treebank dataset sentiment labels for every parse tree constituent, from sentences to to individual words, for 11,855 sentences
Models studied Standard recurrent sequence models with TANH activation functions LSTMs: Long Short Term Memory Bidirectional LSTMs
Visualizations Compositionality Negation Intensification Concessive Salience (contribution of unit to the final composed meaning) Gradient back-propagation The variance of a token from the average word node LSTM-style gates that measure information flow
Local compositionality
t-SNE visualization on representations for negation
Concessive clause composition
Salience Method 1: First derivatives
Salience method 2: Variance of word from sentence mean
Salience method 3: Gate models