T-EBAnO - Text-Explaining BlAck-box mOdels

Explaining predictions of deep learning models for Natural Language Processing tasks.

T-EBAnO is a domain and model specific XAI technique for deep models in the NLP domain.

• Local Explanations: explain single predictions of the model.
• Global Explanations: aggregate multiple local explanations to globally explain the model.

The methodology is model-specific, thus it requires to implement the model_wrapper_interface to adapt for your specific model. Some examples of interface implementations could be found in the model_wrappers folder.

Citation

@article{Ventura2022,
  doi = {10.1007/s10115-022-01690-9},
  url = {https://doi.org/10.1007/s10115-022-01690-9},
  year = {2022},
  month = jun,
  publisher = {Springer Science and Business Media {LLC}},
  author = {Francesco Ventura and Salvatore Greco and Daniele Apiletti and Tania Cerquitelli},
  title = {Trusting deep learning natural-language models via local and global explanations},
  journal = {Knowledge and Information Systems}
}

---

Explainers

LocalExplainer

Produce the local explanations of a set of input texts.

OfflineLocalExplainer

Produces the local explanations of a set of input texts and saves the explanation report persistently in json format. This can be useful also to produce the Global Explanations.

GLobalExplainer

Produces the global explanations by aggregating and analyzing a set of local explanations.

---

Features Extraction

• Multi-layer Word Embedding features:
• Sentence-based features: Extracts a feature and evaluates the impact of each full-sentence in the input text.
• Part-of-speech features: Extracts a feature and evaluates the impact of each part-of-speech (Adjectives, Nouns, etc.)in the input text.

---

Perturbations

• Removal Perturbation: The tokens of each feature are removed from the input text to evaluate their impact in the original prediction of the model.

---

Quantitative Index

• nPIR: normalized Perturbation Influence Relation index

$$nPIR(ci) = softsign(Po) ci: class of interest Po: original probability for the class ci Pf: probability after the perturbation of the feature f for the class ci$$

Experimental Evaluation

---

Human Evaluation

• Survey nPIR correlation with human-judgment: Link to the Survey

---