Learning Explanations from Language Data

David Harbecke, Robert Schwarzenberg, Christoph AltEMNLP 2018

TL;DR

This paper adapts a vision-based neural network explanation method, "Pattern Attribution," to the language domain. They demonstrate its ability to highlight important word contributions for sentiment classification and propose an improved criterion for evaluating explanation methods.

What Problem Does This Paper Address?

Deep neural networks, while powerful, are often black boxes. Understanding why a model makes a particular decision is crucial for debugging, improving performance, and gaining insights into the data itself. Existing explanation methods, particularly those relying solely on model weights, have been shown to be insufficient, even for simple linear models. The authors aim to address this by applying and refining a data-driven explanation technique to the complex domain of natural language processing.

Methods

The core of the paper is the application of "Pattern Attribution," a method originally developed for image classification. This technique aims to learn "patterns" that represent the signal within the input data, rather than just relying on model weights. The authors explain that the original Pattern Attribution method has a quality criterion (Eq. 2) that, while better than weight-based methods, can still be fooled by estimators that heavily scale noise. To address this, they propose a new, more robust quality criterion (Eq. 2') that better distinguishes between signal and noise contributions. For their experiments, they adapt the signal base vector calculation for simple linear and ReLU layers to the NLP context.

Datasets

  • Amazon review polarity data set: A subset of this dataset was used to train a CNN text classifier.
  • No new benchmarks were introduced.

Key Insights and Ideas

  • Pattern Attribution translates well to NLP: The method successfully identifies meaningful contributions of word bigrams to sentiment classification, highlighting that isolated tokens are not typically the focus.
  • Signal vs. Noise is key: The paper emphasizes that effective explanation methods must disentangle the true signal from noise in the input data.
  • Existing criteria can be improved: The authors demonstrate a weakness in the original Pattern Attribution's quality criterion and propose a more robust alternative that better accounts for signal and noise separation.
  • Bigrams are important for sentiment: The experiments show that bigrams with clear positive or negative sentiment contribute significantly to the classification, while stop words and uninformative bigrams have minimal impact.

Key Results

  • The authors successfully applied Pattern Attribution to a CNN text classifier trained on a subset of the Amazon review polarity dataset, achieving an F1 score of 0.875.
  • Qualitative results (shown in Figures 1 and 2) demonstrate that the method highlights relevant bigrams for both positive and negative sentiment classifications, providing visually interpretable explanations.
  • The proposed new quality criterion (Eq. 2') aims to provide a more accurate measure of an explanation method's effectiveness.

Limitations and Future Work

  • Extension to other models: The authors suggest that their method should be extended to other popular NLP models beyond CNNs.
  • Further evaluation of estimators: They leave it to future research to evaluate existing signal estimators using their newly proposed criterion.
  • Deducing estimators: Future work could involve deducing new signal estimators that are specifically designed to perform well under their proposed criterion.

Summary updated

BibTeX
@inproceedings{harbecke2018learning,
  author = {David Harbecke and Robert Schwarzenberg and Christoph Alt},
  title = {{Learning Explanations from Language Data}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2018},
}
Figure