https://bampel.com/tag/model-interpretability/Model InterpretabilityWowchemy (https://wowchemy.com)en-usThu, 29 Oct 2026 00:00:00 +0000https://bampel.com/media/icon_hu_cc9cb5646de94589.pnghttps://bampel.com/tag/model-interpretability/https://bampel.com/workshop_publication/media-bias-blackboxnlp-2026/Thu, 29 Oct 2026 00:00:00 +0000https://bampel.com/workshop_publication/media-bias-blackboxnlp-2026/ <div class="pub-extras"> <div class="pub-plain" role="note" aria-label="Plain-language summary"> <span class="pub-plain-tag">In Plain Terms</span> <p>Bias-detection models can flag a biased news sentence, but that alone doesn&#39;t tell you whether the model is reasoning the way a person would, or whether its stated reasons are what&#39;s actually driving its answer. This paper tests BERT and RoBERTa on media bias detection along three separate axes: how accurate they are, whether their highlighted words match what human experts flagged, and whether those words are truly what the model relies on internally (tested by surgically altering the model&#39;s attention circuits). The three measures often disagree with each other, and pushing a model to highlight the &#39;right&#39; words doesn&#39;t reliably make it rely on them internally. The takeaway for anyone deploying these systems: being accurate, being explainable in a human-relatable way, and being faithful to your own internal reasoning are three different properties, and checking one doesn&#39;t tell you about the others.</p> </div> <div class="pub-extras-grid"> <div class="pub-extras-card"> <h3>Key Contributions</h3> <p class="pub-muted">Key contributions will be added soon.</p> </div> <div class="pub-extras-card"> <h3>Artifacts</h3> <div class="pub-artifacts"> <a class="pub-chip" href="https://bampel.com/uploads/papers/Chen_etal_Ampel_2026_BlackboxNLP_Media-Bias-Explainability.pdf" target="_blank" rel="noopener">PDF</a> </div> </div> <div class="pub-extras-card"> <h3>Related Papers</h3> <ul class="pub-related"> <li> <a href="https://bampel.com/conference_publication/bias-detection-hicss-2026/">A Domain-Adaptive Soft Prompting Framework for Multi-Type Bias Detection in News</a> <span>2026 · HICSS</span> </li> <li> <a href="https://bampel.com/conference_publication/phishing-robustness-amcis-2023/">Benchmarking the Robustness of Phishing Email Detection Systems</a> <span>2023 · AMCIS</span> </li> <li> <a href="https://bampel.com/conference_publication/phishing-spw-2025/">Examining the Robustness of Machine Learning-based Phishing Website Detection: Action-Masked Reinforcement Learning for Automated Red Teaming</a> <span>2025 · IEEE SPW</span> </li> </ul> </div> <div class="pub-extras-card"> <h3>Citation</h3> <div class="pub-citation"> <div class="pub-citation-text">Ting Chen, Chengjun Zhang, Benjamin M. Ampel, &amp; Sagar Samtani (2026). A Multi-Dimensional Evaluation of Explainability in Media Bias Detection. In *Proceedings of BlackboxNLP 2026: The 9th Workshop on Analyzing and Interpreting Neural Networks for NLP*</div> <div class="pub-citation-actions"> <button class="pub-copy" type="button" data-copy-text="Ting&#43;Chen%2C&#43;Chengjun&#43;Zhang%2C&#43;Benjamin&#43;M.&#43;Ampel%2C&#43;%26&#43;Sagar&#43;Samtani&#43;%282026%29.&#43;A&#43;Multi-Dimensional&#43;Evaluation&#43;of&#43;Explainability&#43;in&#43;Media&#43;Bias&#43;Detection.&#43;In&#43;%2AProceedings&#43;of&#43;BlackboxNLP&#43;2026%3A&#43;The&#43;9th&#43;Workshop&#43;on&#43;Analyzing&#43;and&#43;Interpreting&#43;Neural&#43;Networks&#43;for&#43;NLP%2A" data-copy-label="APA">Copy APA</button> <button class="pub-copy" type="button" data-copy-text="Ting&#43;Chen%2C&#43;Chengjun&#43;Zhang%2C&#43;Benjamin&#43;M.&#43;Ampel%2C&#43;and&#43;Sagar&#43;Samtani.&#43;2026.&#43;%22A&#43;Multi-Dimensional&#43;Evaluation&#43;of&#43;Explainability&#43;in&#43;Media&#43;Bias&#43;Detection.%22&#43;In&#43;%2AProceedings&#43;of&#43;BlackboxNLP&#43;2026%3A&#43;The&#43;9th&#43;Workshop&#43;on&#43;Analyzing&#43;and&#43;Interpreting&#43;Neural&#43;Networks&#43;for&#43;NLP%2A" data-copy-label="Chicago">Copy Chicago</button> <button class="pub-copy" type="button" data-copy-text="%40inproceedings%7Bhen2026a%2C%0A&#43;&#43;title%3D%7BA&#43;Multi-Dimensional&#43;Evaluation&#43;of&#43;Explainability&#43;in&#43;Media&#43;Bias&#43;Detection%7D%2C%0A&#43;&#43;author%3D%7BTing&#43;Chen&#43;and&#43;Chengjun&#43;Zhang&#43;and&#43;Benjamin&#43;M.&#43;Ampel&#43;and&#43;Sagar&#43;Samtani%7D%2C%0A&#43;&#43;booktitle%3D%7BIn&#43;%2AProceedings&#43;of&#43;BlackboxNLP&#43;2026%3A&#43;The&#43;9th&#43;Workshop&#43;on&#43;Analyzing&#43;and&#43;Interpreting&#43;Neural&#43;Networks&#43;for&#43;NLP%2A%7D%2C%0A&#43;&#43;year%3D%7B2026%7D%0A%7D" data-copy-label="BibTeX">Copy BibTeX</button> </div> </div> </div> </div> </div> <style> .pub-extras { margin-top: 2rem; } .pub-plain { position:relative; border:1px solid rgba(0,229,255,0.35); border-left:4px solid var(--clr-cyan, #00e5ff); border-radius:14px; padding:16px 18px 14px; margin-bottom:18px; background:linear-gradient(180deg, rgba(0,229,255,0.10), rgba(0,229,255,0.04)); } .pub-plain-tag { display:inline-block; font-size:0.66rem; font-weight:700; text-transform:uppercase; letter-spacing:0.12em; color:var(--clr-cyan, #00e5ff); margin-bottom:6px; } .pub-plain p { margin:0; font-size:0.95rem; line-height:1.55; color:var(--text, #c9d1d9); } body:not(.dark) .pub-plain { background:rgba(0,150,180,0.06); border-color:rgba(0,120,150,0.35); } body:not(.dark) .pub-plain-tag { color:#0b6b80; } body:not(.dark) .pub-plain p { color:#1a2a33; } .pub-extras-grid { display:grid; grid-template-columns: repeat(auto-fit, minmax(240px, 1fr)); gap:16px; } .pub-extras-card { border:1px solid var(--card-border, #333); border-radius:14px; padding:16px; background:var(--card-bg, rgba(0,0,0,0.25)); box-shadow:0 8px 20px rgba(0,0,0,0.12); } .pub-extras-card h3 { margin:0 0 10px; font-size:0.9rem; text-transform:uppercase; letter-spacing:0.08em; color:var(--clr-green); } .pub-extras-card ul { margin:0; padding-left:18px; } .pub-extras-card li { margin-bottom:6px; font-size:0.82rem; line-height:1.4; } .pub-muted { font-size:0.8rem; opacity:0.7; } .pub-artifacts { display:flex; flex-wrap:wrap; gap:8px; } .pub-chip { font-size:0.7rem; text-transform:uppercase; letter-spacing:0.08em; padding:6px 10px; border-radius:999px; border:1px solid rgba(0,229,255,0.35); color:#bde8ff; background:rgba(0,229,255,0.1); text-decoration:none; } .pub-related { list-style:none; padding:0; margin:0; display:flex; flex-direction:column; gap:8px; } .pub-related li { display:flex; flex-direction:column; gap:2px; } .pub-related a { color:var(--clr-cyan); text-decoration:none; font-size:0.82rem; } .pub-related span { font-size:0.72rem; opacity:0.7; } .pub-citation { display:flex; flex-direction:column; gap:10px; } .pub-citation-text { font-size:0.8rem; line-height:1.4; background:rgba(0,0,0,0.35); border-radius:10px; padding:10px; border:1px solid rgba(0,255,65,0.2); } .pub-citation-actions { display:flex; flex-wrap:wrap; gap:8px; } .pub-copy { align-self:flex-start; border-radius:999px; border:1px solid rgba(0,255,65,0.4); color:var(--clr-green); background:rgba(0,255,65,0.12); padding:6px 12px; font-size:0.7rem; text-transform:uppercase; letter-spacing:0.08em; cursor:pointer; } .pub-copy.copied { color:#111; background:var(--clr-green); } body:not(.dark) .pub-extras-card { background:rgba(255,255,255,0.92); } body:not(.dark) .pub-extras-card h3 { color:#0b7a2f; } body:not(.dark) .pub-citation-text { background:rgba(255,255,255,0.9); } </style>