ProtoTEx: Explaining Model Decisions with Prototype Tensors
This addresses the need for faithful explanations in NLP models, particularly for non-experts in tasks like propaganda detection, though it is incremental as it builds on existing prototype network methods.
The paper tackles the problem of making NLP classification models more interpretable by introducing ProtoTEx, a white-box architecture that uses prototype tensors to explain decisions based on training examples, achieving accuracy matching BART-large and exceeding BERT-large on a propaganda detection task.
We present ProtoTEx, a novel white-box NLP classification architecture based on prototype networks. ProtoTEx faithfully explains model decisions based on prototype tensors that encode latent clusters of training examples. At inference time, classification decisions are based on the distances between the input text and the prototype tensors, explained via the training examples most similar to the most influential prototypes. We also describe a novel interleaved training algorithm that effectively handles classes characterized by the absence of indicative features. On a propaganda detection task, ProtoTEx accuracy matches BART-large and exceeds BERT-large with the added benefit of providing faithful explanations. A user study also shows that prototype-based explanations help non-experts to better recognize propaganda in online news.