Analysing Neural Language Models: Contextual Decomposition Reveals Default Reasoning in Number and Gender Assignment
This provides insight into model interpretability for researchers in NLP, but it is incremental as it builds on existing decomposition methods.
The authors tackled the problem of understanding how recurrent neural language models process grammatical phenomena by proposing a generalization of Contextual Decomposition (GCD) to analyze predictions. They discovered that the models strongly rely on a default reasoning effect for tasks like syntactic agreement and co-reference resolution.
Extensive research has recently shown that recurrent neural language models are able to process a wide range of grammatical phenomena. How these models are able to perform these remarkable feats so well, however, is still an open question. To gain more insight into what information LSTMs base their decisions on, we propose a generalisation of Contextual Decomposition (GCD). In particular, this setup enables us to accurately distil which part of a prediction stems from semantic heuristics, which part truly emanates from syntactic cues and which part arise from the model biases themselves instead. We investigate this technique on tasks pertaining to syntactic agreement and co-reference resolution and discover that the model strongly relies on a default reasoning effect to perform these tasks.