Argumentative explanations for pattern-based text classifiers

Recent works in Explainable AI mostly address the transparency issue of black-box models or create explanations for any kind (i.e., they are model-agnostic), while leaving interpretable largely underexplored. In this paper, we fill gap by focusing on a specific model, namely pattern-based logistic regression (PLR) binary text classification. We do so because, albeit interpretable, PLR is challenging when it comes to explanations. particular, found that standard way extract from model does not consider relations among features, making hardly plausible humans. Hence, propose AXPLR, novel explanation method using (forms of) computational argumentation generate (for outputs computed PLR) which unearth agreements and disagreements features. Specifically, use as follows: see features (patterns) arguments form quantified bipolar frameworks (QBAFs) attacks supports between based specificity arguments; understand gradual semantics these QBAFs, used determine arguments’ dialectic strength; study properties QBAFs context our argumentative re-interpretation PLR, sanctioning its suitability explanatory purposes. then show how intuitive constructed QBAFs. Finally, conduct an empirical evaluation two experiments human-AI collaboration demonstrate advantages resulting AXPLR method.