| <!DOCTYPE html> |
| <html> |
| <head> |
| <meta charset="UTF-8"> |
| <title>Text-PRAS Final Baseline Adjusted Review Version</title> |
| <style> |
| body { font-family: 'Times New Roman', serif; font-size: 12pt; line-height: 1.35; margin: 1in; } |
| h1 { font-size: 18pt; text-align: center; } |
| h2 { font-size: 15pt; margin-top: 18pt; } |
| h3 { font-size: 13pt; margin-top: 14pt; } |
| table { border-collapse: collapse; width: 100%; margin: 12pt 0; font-size: 10pt; } |
| th, td { border: 1px solid #444; padding: 5px; vertical-align: top; } |
| th { background: #eeeeee; } |
| .caption { font-style: italic; text-align: center; margin-bottom: 12pt; } |
| .eq { font-family: 'Cambria Math', 'Times New Roman', serif; margin-left: 24pt; } |
| .note { color: #444; font-style: italic; } |
| </style> |
| </head> |
| <body> |
|
|
| <h1>Text-PRAS: An Explanation-Alignment Audit Framework for Arabic Stance Detection</h1> |
| <p style="text-align:center;"><b>Anonymous Authors</b><br/>Anonymous Affiliation<br/>anonymous@example.com</p> |
|
|
| <h2>Abstract</h2> |
| <p>We introduce <b>Text-PRAS</b>, an attribution-mass audit framework for Arabic target-conditioned stance detection. Text-PRAS quantifies how post-hoc attribution mass is distributed over predefined stance-evidence and shortcut/noise token categories. On the official public Mawqif-v2 train/dev split using MARBERT, raw Text-PRAS scores are positive across IG, GradShap, and LIME. However, within-example shuffled attribution baselines are consistently higher, yielding negative baseline adjusted deltas. This shows that raw positive attribution-mass scores can overstate explanation alignment unless controlled against Atlas/token coverage. Text-PRAS is therefore presented as an audit signal, not as proof of true reasoning or explanation faithfulness.</p> |
|
|
| <h2>1. Introduction</h2> |
| <p>Arabic stance detection has important applications in social media analytics, public opinion monitoring, and misinformation analysis. However, high classification scores do not establish whether models rely on stance-relevant textual evidence or on shortcut cues such as hashtags, URLs, mentions, repeated slogans, or boilerplate text. We introduce TextEvidenceAtlas and Text-PRAS as an explanation-alignment audit framework for Arabic target-conditioned stance detection. The goal is not leaderboard optimization, but a reproducible audit signal that complements conventional stance metrics.</p> |
|
|
| <h3>Why Arabic requires a dedicated audit</h3> |
| <p>Arabic social-media text presents attribution challenges that are not adequately handled by directly translating English rationale or shortcut taxonomies. Arabic is morphologically rich: clitics attach to words, negation and prepositions may be short or attached, and subword tokenization can fragment stance-bearing words. Dialectal and orthographic variation produce many surface forms for the same stance cue. Arabic-script hashtags often conflate ordinary words, target mentions, campaign slogans, and stance markers in a single token. Sarcasm, contrast, and quotation patterns in right-to-left social media text may invert the stance implied by local sentiment words. TextEvidenceAtlas categories therefore explicitly include Arabic stance markers, negation/modality cues, contrast cues, target anchors, Arabic sentiment cues, and Arabic Twitter artifacts such as hashtags, mentions, repeated emojis, repeated slogans, and boilerplate phrases. This makes the framework task-general within Arabic target-conditioned stance detection, but not universal across Arabic NLP tasks.</p> |
|
|
| <h3>Contributions</h3> |
| <p>We make three bounded contributions. First, we introduce TextEvidenceAtlas, a transparent rule-based atlas for Arabic target-conditioned stance detection that distinguishes stance evidence from social-media shortcuts. Second, we define Text-PRAS, an attribution-mass audit score that measures evidence mass minus shortcut mass and explicitly report coverage, sensitivity diagnostics, overlap diagnostics, and a within-example random baseline. Third, we evaluate the audit on MARBERT over the Mawqif-v2 official public train/dev split using three attribution methods, showing both method agreement (IG/GradShap) and method dependence (LIME), while avoiding claims of reasoning or universal validity.</p> |
|
|
| <h3>Scope</h3> |
| <p>The unseen test is pending official release. We do not claim final unseen performance, model understanding, causality, or proof of reasoning. Text-PRAS is an explanation-alignment audit signal.</p> |
|
|
| <h2>2. Related Work</h2> |
| <p><b>Arabic stance detection and Arabic social-media modeling.</b> Arabic stance detection has been studied in fact-checking, claim verification, and social-media settings. AraStance provides a multi-country and multi-domain Arabic stance dataset for fact checking, while Arabic claim-verification work studies stance prediction under limited contextual evidence. Mawqif-v2 extends this line by focusing on target-conditioned Arabic tweets with Favor, Against, and None labels. Our work differs from modeling-centered studies: we do not aim to improve leaderboard performance, but to audit whether attribution mass for a trained Arabic stance model aligns with predefined stance-relevant textual evidence rather than social-media shortcuts.</p> |
|
|
| <p><b>Interpretability and attribution in NLP.</b> Post-hoc explanation methods such as LIME, Integrated Gradients, and SHAP-style methods provide token-level importance estimates, but their outputs are not ground truth explanations. Prior work emphasizes that explanation faithfulness must be defined carefully rather than assumed from plausibility alone. We follow this cautionary line: Text-PRAS is not a claim that the model understands stance; it is an explanation-alignment audit signal measuring whether attribution mass falls on atlas-defined categories.</p> |
|
|
| <p><b>Shortcut learning and annotation artifacts.</b> NLP models often achieve high task performance by exploiting artifacts or shallow heuristics. Our shortcut categories, such as hashtags, URLs, mentions, repeated emojis, repeated slogans, and boilerplate phrases, are motivated by this literature but adapted to Arabic social-media stance detection.</p> |
|
|
| <h2>3. TextEvidenceAtlas</h2> |
| <p>TextEvidenceAtlas is a predefined rule-based atlas. It is task-general within Arabic target-conditioned stance detection, not a universal Arabic NLP metric.</p> |
|
|
| <table> |
| <tr><th>Category</th><th>Weight</th><th>Arabic examples</th></tr> |
| <tr><td>Explicit stance markers</td><td>+1.00</td><td dir="rtl">أدعم، أؤيد، أرفض، ضد، مع</td></tr> |
| <tr><td>Target-conditioned reason</td><td>+0.85</td><td dir="rtl">ضروري، أساسي، حيوي</td></tr> |
| <tr><td>Target anchor/mention</td><td>+0.60</td><td dir="rtl">اللقاح، التمكين، target name</td></tr> |
| <tr><td>Negation/modality</td><td>+0.50</td><td dir="rtl">لا، ليس، غير، لن</td></tr> |
| <tr><td>Contrast/concession</td><td>+0.40</td><td dir="rtl">لكن، بالرغم من، على الرغم من</td></tr> |
| <tr><td>General sentiment</td><td>+0.30</td><td dir="rtl">رائع، سيء، ممتاز، فاشل</td></tr> |
| <tr><td>Function/stopword</td><td>0.00</td><td dir="rtl">في، من، إلى، و، أن</td></tr> |
| </table> |
| <p class="caption">Table 1. Evidence categories in TextEvidenceAtlas v1. Weights are predefined audit choices, not calibrated human-validated importance values.</p> |
|
|
| <table> |
| <tr><th>Category</th><th>Weight</th><th>Examples</th></tr> |
| <tr><td>URLs</td><td>-0.50</td><td>http://, URL</td></tr> |
| <tr><td>Mentions</td><td>-0.50</td><td>@user</td></tr> |
| <tr><td>Hashtags</td><td>-0.50</td><td>#tag</td></tr> |
| <tr><td>Repeated emojis</td><td>-0.50</td><td>repeated emoji sequences</td></tr> |
| <tr><td>Target-only reliance</td><td>-0.50</td><td>tweet equals target only</td></tr> |
| <tr><td>Repeated slogans</td><td>-0.50</td><td>repeated slogan phrases</td></tr> |
| <tr><td>Boilerplate phrases</td><td>-0.50</td><td>copy/paste text</td></tr> |
| </table> |
| <p class="caption">Table 2. Shortcut categories in TextEvidenceAtlas v1.</p> |
|
|
| <p><b>Overlap caveat.</b> Atlas rules may assign overlapping evidence and shortcut labels to some tokens, especially target-related or hashtag-like tokens. We therefore report overlap diagnostics and rely on the canonical attribution-mass accounting used by Text-PRAS.</p> |
|
|
| <table> |
| <tr><th>Method</th><th>Records</th><th>Records with overlap</th><th>Overlap token assignments</th></tr> |
| <tr><td>IG</td><td>619</td><td>153</td><td>342</td></tr> |
| <tr><td>GradShap</td><td>619</td><td>153</td><td>342</td></tr> |
| <tr><td>LIME</td><td>619</td><td>161</td><td>325</td></tr> |
| </table> |
| <p class="caption">Table 3. Evidence/shortcut overlap diagnostics from strict Magnolia verification.</p> |
|
|
| <h2>4. Text-PRAS</h2> |
| <p>For token attribution score <i>a<sub>i</sub></i>:</p> |
| <p class="eq">Text-PRAS = EvidenceMass − ShortcutMass</p> |
| <p class="eq">EvidenceMass = Σ<sub>i</sub> |a<sub>i</sub>| I[t<sub>i</sub> ∈ E] / Σ<sub>i</sub> |a<sub>i</sub>|</p> |
| <p class="eq">ShortcutMass = Σ<sub>i</sub> |a<sub>i</sub>| I[t<sub>i</sub> ∈ S] / Σ<sub>i</sub> |a<sub>i</sub>|</p> |
| <p>Text-PRAS lies in [-1, 1]. We report raw Text-PRAS, Atlas v1 raw-score sensitivity, and within-example shuffle baselines.</p> |
|
|
| <h3>Within-example random attribution baseline</h3> |
| <p>Because Atlas coverage influences raw Text-PRAS, positive raw scores alone are insufficient. We use a within-example shuffle baseline: for each example, attribution magnitudes are shuffled only within that same example while preserving the same text, target, labels, tokens, and Atlas category positions. We compute delta = observed Text-PRAS − mean shuffled Text-PRAS, and bootstrap over examples.</p> |
|
|
| <h2>5. Experimental Setup</h2> |
| <table> |
| <tr><th>Split</th><th>Size</th><th>Status</th><th>Notes</th></tr> |
| <tr><td>Train</td><td>3,502</td><td>used</td><td>official public train</td></tr> |
| <tr><td>Dev</td><td>619</td><td>used</td><td>official public dev</td></tr> |
| <tr><td>Unseen test</td><td>996</td><td>pending</td><td>not evaluated here</td></tr> |
| </table> |
| <p class="caption">Table 4. Mawqif-v2 official public data setup.</p> |
|
|
| <p>We fine-tune UBC-NLP/MARBERT for 3-class sequence classification using max length 128, batch size 8, 3 epochs, learning rate 2×10<sup>−5</sup>, seed 42, AdamW, weight decay 0.01, and early stopping on dev Favg2. The selected input format is Format A: <code>Target: <target> [SEP] Tweet: <tweet></code>.</p> |
|
|
| <h2>6. Results</h2> |
| <h3>Raw Text-PRAS and shuffle baseline</h3> |
| <table> |
| <tr><th>Method</th><th>Observed</th><th>Shuffle</th><th>Delta</th><th>95% CI</th></tr> |
| <tr><td>IG</td><td>0.1075</td><td>0.1204</td><td>-0.0129</td><td>[-0.0179, -0.0077]</td></tr> |
| <tr><td>GradShap</td><td>0.1074</td><td>0.1204</td><td>-0.0130</td><td>[-0.0185, -0.0076]</td></tr> |
| <tr><td>LIME</td><td>0.1265</td><td>0.1493</td><td>-0.0229</td><td>[-0.0315, -0.0149]</td></tr> |
| </table> |
| <p class="caption">Table 5. Within-example shuffled baseline. The key quantity is delta and its confidence interval.</p> |
|
|
| <p>Raw Text-PRAS scores are positive, but the within-example shuffled baselines are consistently higher. Therefore, baseline adjusted deltas are negative for all three methods. This means raw positive Text-PRAS would have overstated explanation alignment if not controlled against Atlas/token coverage.</p> |
|
|
| <h3>Strict attribution-mass coverage</h3> |
| <table> |
| <tr><th>Method</th><th>Evidence mass</th><th>Shortcut mass</th><th>Neutral mass</th></tr> |
| <tr><td>IG</td><td>16.25%</td><td>5.50%</td><td>78.25%</td></tr> |
| <tr><td>GradShap</td><td>16.28%</td><td>5.54%</td><td>78.17%</td></tr> |
| <tr><td>LIME</td><td>20.67%</td><td>8.03%</td><td>71.30%</td></tr> |
| </table> |
| <p class="caption">Table 6. Strict attribution-mass coverage from canonical Magnolia Text-PRAS files.</p> |
|
|
| <h3>Strict tokenization-aware Atlas coverage</h3> |
| <table> |
| <tr><th>Source</th><th>Token unit</th><th>Total tokens</th><th>Evidence tokens</th><th>Shortcut tokens</th><th>Neutral tokens</th><th>Overlap tokens</th></tr> |
| <tr><td>IG</td><td>MARBERT subwords</td><td>30,728</td><td>4,861</td><td>1,324</td><td>24,201</td><td>342</td></tr> |
| <tr><td>GradShap</td><td>MARBERT subwords</td><td>30,728</td><td>4,861</td><td>1,324</td><td>24,201</td><td>342</td></tr> |
| <tr><td>LIME</td><td>LIME words</td><td>20,529</td><td>3,981</td><td>932</td><td>15,291</td><td>325</td></tr> |
| </table> |
| <p class="caption">Table 7. Strict tokenization-aware Atlas coverage. Tokenization units differ across attribution methods and should not be mixed.</p> |
|
|
| <h3>Atlas v1 sensitivity</h3> |
| <p>Atlas v1 sensitivity should be interpreted as raw-score sensitivity analysis, not random adjusted evidence alignment. The raw-score sensitivity table is valid only as a robustness check over scoring modes; it does not show above-random evidence alignment.</p> |
|
|
| <h2>7. Discussion</h2> |
| <p><b>Raw positivity versus baseline adjusted alignment.</b> Although raw Text-PRAS is positive for all methods, within-example shuffled baselines are consistently higher, with negative deltas whose confidence intervals lie below zero. This indicates that attribution mass is not concentrated on evidence tokens above what Atlas/token coverage alone would predict under random attribution. We therefore treat Text-PRAS as a diagnostic audit signal rather than evidence of above-random evidence alignment.</p> |
|
|
| <p><b>Evidence alignment is not sufficient for correctness.</b> Text-PRAS measures attribution alignment with atlas-defined evidence categories, not correctness, compositional understanding, or causal reliance. An incorrect prediction can still focus on stance-bearing words, target mentions, or negation cues while misreading direction, sarcasm, scope, or target relation. Text-PRAS should be used alongside, not instead of, task metrics.</p> |
|
|
| <h2>8. Conclusion</h2> |
| <p>We introduced Text-PRAS, an attribution-mass audit signal for Arabic stance detection. The finding that observed Text-PRAS falls below within-example shuffle baselines is itself a meaningful audit result: it demonstrates that the framework can distinguish between raw positivity driven by Atlas/token coverage and genuine above-random evidence alignment. Text-PRAS is therefore useful as an audit signal precisely because it prevents overclaiming from raw positive attribution-mass scores.</p> |
|
|
| <h2>Limitations</h2> |
| <ol> |
| <li>Results are bounded to Mawqif-v2, MARBERT, and the official public dev split.</li> |
| <li>The unseen test is pending official release; no final unseen result is claimed.</li> |
| <li>IG, GradShap, and LIME are post-hoc approximations; none is ground truth.</li> |
| <li>The atlas is rule-based and may miss morphology, dialect, sarcasm scope, and emerging slang.</li> |
| <li>Atlas rules may produce overlapping evidence and shortcut labels; overlap is reported diagnostically.</li> |
| <li>Positive raw Text-PRAS does not imply correctness, causality, true reasoning, or explanation faithfulness.</li> |
| </ol> |
|
|
| <h2>References</h2> |
| <p>References should be compiled from <code>references_expanded.bib</code> in the ACL LaTeX version.</p> |
|
|
| </body> |
| </html> |
|
|
