Evaluating neural network explanation methods using hybrid documents and morphological prediction
share
We propose two novel paradigms for evaluating neural network explanations in NLP. The first paradigm works on hybrid documents, the second exploits morphosyntactic agreements. Neither paradigm requires manual annotations; instead, a relevance ground truth is generated automatically. In our experiments, successful explanations for Long Short Term Memory networks (LSTMs) were produced by a decomposition of memory cells (Murdoch & Szlam, 2017), while for convolutional neural networks, a gradient-based method by (Denil et al., 2014) works well. We also introduce LIMSSE, a substring-based extension of LIME (Ribeiro et al., 2016) that produces the most successful explanations in the hybrid document experiment.
READ FULL TEXT