Datasets:

Dr-RazanAlsulieman
/

XAI-Safety-Judge-NLP-v4-HPC

Tags:

Dr-RazanAlsulieman commited on 21 days ago

Commit

f4d9bc9

verified ·

1 Parent(s): bbc2d51

Add polished full Word manuscript with expanded references

Browse files

Files changed (1) hide show

TextPRAS_polished_manuscript_with_references_full.doc +123 -0

TextPRAS_polished_manuscript_with_references_full.doc ADDED Viewed

	@@ -0,0 +1,123 @@

+<!DOCTYPE html><html><head><meta charset="UTF-8"><title>Text-PRAS Polished Manuscript</title>
+<style>body{font-family:'Times New Roman',serif;font-size:12pt;line-height:1.35;margin:1in} h1{text-align:center;font-size:18pt} h2{font-size:15pt;margin-top:18pt} h3{font-size:13pt;margin-top:14pt} table{border-collapse:collapse;width:100%;font-size:10pt;margin:12pt 0} th,td{border:1px solid #444;padding:5px;vertical-align:top} th{background:#eee} code{font-family:'Courier New',monospace}</style></head><body>
+<h1>Text-PRAS: A Baseline-Adjusted Attribution-Mass Audit Framework for Arabic Stance Detection</h1>
+<p>Anonymous authors</p>
+<h2>Abstract</h2>
+<p>We introduce Text-PRAS, 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. The contribution is a task-conditioned audit protocol that combines an Arabic stance-evidence atlas, post-hoc attribution mass, tokenization-aware coverage diagnostics, overlap diagnostics, and within-example shuffled baselines. Text-PRAS is therefore presented as an audit signal, not proof of true reasoning or explanation faithfulness.</p>
+<h2>1. Introduction</h2>
+<p>Arabic stance detection asks whether a text expresses Favor, Against, or None toward a specified target. Stance detection has been studied in shared-task and social-media settings such as SemEval stance detection and Arabic target-specific stance datasets (Mohammad et al. (2016); Mawqif dataset (KFUPM); StanceEval (2026); Baly et al. / AraStance (2021); Stance Prediction and Claim Verification (2020)). In social-media Arabic, high classification performance does not by itself show whether a model relies on stance-relevant evidence or shortcut cues such as URLs, hashtags, target-only phrases, repeated slogans, or boilerplate.</p>
+<p>We introduce Text-PRAS as a task-conditioned audit framework for Arabic stance detection. Text-PRAS does not introduce a new attribution method and does not claim to prove reasoning. Instead, it audits post-hoc attribution maps by measuring how much absolute attribution mass falls on predefined stance-evidence categories versus shortcut/noise categories. The audit is explicitly baseline-adjusted: it compares observed Text-PRAS to within-example shuffled attribution baselines that preserve each example&#x27;s attribution magnitudes and Atlas category layout.</p>
+<p>The resulting finding is conservative. Raw Text-PRAS values are positive for IG, GradShap, and LIME, but the within-example shuffled baselines are consistently higher. This means that raw positive attribution-mass scores would have overstated explanation alignment if not controlled against Atlas/token coverage. We treat this as a meaningful audit result rather than a failure of the framework: the framework detects that positive raw attribution mass does not necessarily imply evidence alignment above the within-example shuffled baseline.</p>
+<p>Our contributions are threefold. First, we define TextEvidenceAtlas, a transparent Arabic stance-evidence and shortcut/noise atlas for target-conditioned stance detection. Second, we define Text-PRAS as an attribution-mass audit signal with tokenization-aware coverage, overlap diagnostics, and within-example shuffled baselines. Third, we apply the audit to a frozen MARBERT stance model on the official public Mawqif-v2 train/dev split, reporting both raw and baseline-adjusted results without claiming explanation faithfulness, causality, model understanding, or official unseen-test performance.</p>
+<h2>2. Related Work</h2>
+<h3>2.1 Arabic stance detection and Mawqif/StanceEval</h3>
+<p>Target-conditioned stance detection has been studied in English and multilingual settings, including SemEval-style stance tasks (Mohammad et al. (2016)). Arabic stance detection includes target-specific Arabic stance datasets, fact-checking-oriented stance resources, and Arabic social-media datasets (Mawqif dataset (KFUPM); StanceEval (2026); Baly et al. / AraStance (2021); Stance Prediction and Claim Verification (2020); Baly et al. (2018); Alturayeif et al. (2022)). Mawqif-v2 and StanceEval-style data are relevant because they provide an Arabic target-conditioned stance setup over social-media text and evaluate performance with stance-specific metrics such as Favg2. Our work uses the official public Mawqif-v2 train/dev split and treats the official unseen test as pending official release/evaluation; no official unseen-test result is claimed. We do not substitute external Arabic stance data for the official unseen split, because doing so would conflate unseen-target generalization with dataset shift and annotation-protocol differences.</p>
+<p>Arabic social-media modeling motivates the use of Arabic pretrained models. MARBERT/ARBERT, AraBERT, QARiB, and AraELECTRA show the importance of Arabic-specific and social-media-aware language modeling (Abdul-Mageed et al. (2021); Antoun et al. (2020); Abdelali et al. (2021); AraELECTRA (2020)). We use MARBERT because it is designed for Arabic social-media text.</p>
+<h3>2.2 Rationale evaluation and evidence alignment in NLP</h3>
+<p>Rationale and evidence evaluation in NLP includes work on annotator rationales, neural rationales, natural-language explanations, and ERASER-style benchmark evaluation (Zaidan et al. (2007); Lei et al. (2016); Camburu et al. (2018); DeYoung et al. (2020); Jain et al. (2020)). Text-PRAS builds on the broad goal of evidence-oriented evaluation, but differs in what is audited. It does not use human rationale annotations, and it does not evaluate whether a generated rationale is faithful. Instead, it measures post-hoc attribution mass over predefined task-conditioned audit categories: stance-evidence tokens and shortcut/noise tokens. This makes Text-PRAS a lightweight audit protocol rather than a human-rationale benchmark.</p>
+<h3>2.3 Post-hoc attribution methods in NLP</h3>
+<p>Post-hoc attribution methods such as LIME, Integrated Gradients, SHAP-style methods, and Captum GradientShap provide token-level importance estimates for model decisions (Ribeiro et al. (2016); Sundararajan et al. (2017); Lundberg and Lee (2017); Captum (2024)). These methods are useful for auditing, but they are approximations and should not be treated as ground-truth explanations. Work on saliency sanity checks, interpretability benchmarks, attention-based explanations, and user-centered explanation evaluation further cautions against equating plausible explanation displays with faithful model reasoning (Adebayo et al. (2018); Hooker et al. (2019); Jain and Wallace (2019); Wiegreffe and Pinter (2019); Hase and Bansal (2020); Pruthi et al. (2020)).</p>
+<h3>2.4 Shortcut learning, spurious cues, and faithfulness limits</h3>
+<p>Prior work on annotation artifacts, hypothesis-only baselines, syntactic heuristics, counterfactually augmented data, and shortcut learning shows that models can achieve strong task metrics while relying on shallow or spurious cues (Gururangan et al. (2018); Poliak et al. (2018); McCoy et al. (2019); Kaushik et al. (2020); Geirhos et al. (2020)). Faithfulness-focused interpretability work similarly warns that plausible explanations are not necessarily faithful explanations (Jacovi and Goldberg (2020); Adebayo et al. (2018); Pruthi et al. (2020)). Text-PRAS builds on prior work in rationale evaluation, post-hoc attribution, and shortcut analysis, but adapts these ideas into a task-conditioned Arabic stance-detection audit framework with Atlas-based evidence/shortcut categories and within-example shuffled baselines.</p>
+<h2>3. TextEvidenceAtlas</h2>
+<p>TextEvidenceAtlas is a rule-based, task-conditioned atlas for Arabic stance detection. It marks tokens as potential stance evidence, shortcut/noise, or neutral according to predefined rules. Evidence categories include explicit stance markers, target-conditioned reasons, target anchors, negation/modality, contrast/concession, and sentiment cues. Shortcut/noise categories include URLs, mentions, hashtags, repeated emojis, target-only reliance, repeated slogans, and boilerplate.</p>
+<p>The Atlas includes predefined weights for raw-score sensitivity analysis. These weights are editorial audit choices, not learned parameters and not empirically validated importance values. The canonical baseline-adjusted result does not depend on interpreting these weights as calibrated evidence strength. The weighted Atlas is used only as a raw-score sensitivity variant. It should not be interpreted as calibrated evidence strength or as the primary baseline-adjusted alignment result.</p>
+<p>Atlas rules may assign overlapping evidence and shortcut labels to some tokens, especially target-related or hashtag-like tokens. Overlap is reported as a diagnostic property of the rule labels. The reported Text-PRAS values follow the canonical implementation used in the released audit files; overlap counts are reported to make this ambiguity transparent rather than hidden.</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>Coverage is tokenization-aware. IG and GradShap use MARBERT subwords, while LIME uses word-like tokens. Therefore, coverage values must not be mixed across tokenization units.</p>
+<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>
+<h2>4. Text-PRAS Method</h2>
+<p>For token attribution score a_i, Text-PRAS is defined as:</p>
+<p>Text-PRAS = EvidenceMass - ShortcutMass</p>
+<p>EvidenceMass is the sum of absolute attribution mass over Atlas evidence categories divided by total absolute attribution mass. ShortcutMass is the analogous quantity over shortcut/noise categories. Text-PRAS is an attribution-mass audit signal, not proof of model understanding, causality, or explanation faithfulness.</p>
+<p>The within-example shuffled baseline is computed as follows. For each example, we keep the text, target, labels, tokens, Atlas categories, and attribution magnitudes fixed. We shuffle attribution magnitudes only within that example across token/category positions. For each example, we compute delta = observed Text-PRAS minus mean shuffled Text-PRAS. We then bootstrap over examples and report a 95% percentile confidence interval for the mean delta.</p>
+<h2>5. Experimental Setup</h2>
+<p>We use the official public Mawqif-v2 train/dev split: 3,502 training rows and 619 development rows. Labels are Favor, Against, and None. The model is UBC-NLP/MARBERT (Abdul-Mageed et al. (2021)). The input format is target-conditioned: <code>Target: &lt;target&gt; [SEP] Tweet: &lt;tweet&gt;</code>.</p>
+<p>We compare three prompt formats on the official development set. Format A is selected because it achieves the highest Favg2, the primary ranking metric, even though other metrics may slightly favor other formats.</p>
+<table>
+<tr><th>Format</th><th>Accuracy</th><th>Macro-F1 / Favg3</th><th>Favg2</th><th>Favor F1</th><th>Against F1</th><th>None F1</th></tr>
+<tr><td>A</td><td>0.8045</td><td>0.6956</td><td>0.8191</td><td>0.8748</td><td>0.7634</td><td>0.4486</td></tr>
+<tr><td>B</td><td>0.8013</td><td>0.6933</td><td>0.8156</td><td>0.8727</td><td>0.7586</td><td>0.4486</td></tr>
+<tr><td>C</td><td>0.8078</td><td>0.7004</td><td>0.8147</td><td>0.8795</td><td>0.7500</td><td>0.4717</td></tr>
+</table>
+<p>Attributions are extracted for the predicted class using IG, GradShap, and LIME. Absolute attribution mass is used in all Text-PRAS accounting. Atlas v1 sensitivity is reported as raw-score sensitivity analysis, not random-adjusted evidence alignment.</p>
+<h2>6. Results</h2>
+<h3>6.1 Raw Text-PRAS and within-example shuffled 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>Raw Text-PRAS scores are positive for all three attribution methods. However, within-example shuffled baselines are consistently higher. The baseline-adjusted deltas are negative and their confidence intervals lie below zero. Thus, raw positive attribution-mass scores are not above what Atlas/token coverage predicts under random attribution.</p>
+<h3>6.2 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>These strict attribution-mass coverage values are computed from the canonical Text-PRAS audit outputs.</p>
+<h3>6.3 Tokenization-aware 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>Tokenization-aware coverage shows that coverage diagnostics depend strongly on attribution token units. For this reason, coverage should be reported with explicit tokenization unit and source.</p>
+<h3>6.4 Atlas v1 sensitivity</h3>
+<p>Atlas v1 sensitivity results remain valid as raw-score sensitivity analysis, not random-adjusted evidence alignment. They should not be used to claim evidence alignment above the within-example shuffled baseline.</p>
+<h2>7. Discussion</h2>
+<p>Text-PRAS did not validate the model; it audited it. The audit revealed that positive raw attribution-mass scores can disappear after controlling for Atlas/token coverage through a within-example shuffled baseline. This is the central finding of the paper. Negative baseline-adjusted deltas are a meaningful audit finding because they prevent overclaiming from raw positive attribution scores.</p>
+<p>Evidence alignment is also not sufficient for correctness. A model can attend to stance-bearing words, target mentions, or negation cues while still misreading direction, sarcasm, scope, or target relation. Text-PRAS should therefore be used alongside task metrics, not instead of them. LIME having higher raw evidence mass should not be interpreted as LIME being better or the model being better, because the baseline-adjusted result remains negative.</p>
+<h2>8. Limitations</h2>
+<p>The results are bounded to Mawqif-v2, MARBERT, and the official public development set. No official unseen-test result is claimed. IG, GradShap, and LIME are post-hoc approximations and are not ground-truth explanations. TextEvidenceAtlas is rule-based and task-conditioned; a new task-specific Atlas would be required for other Arabic NLP tasks. Atlas weights are predefined editorial audit weights, not learned or calibrated parameters. Overlap exists between evidence and shortcut labels. Tokenization differences affect coverage diagnostics. Positive raw Text-PRAS does not imply correctness, causality, true reasoning, or explanation faithfulness.</p>
+<h2>9. Conclusion</h2>
+<p>Text-PRAS is a baseline-adjusted attribution-mass audit framework for Arabic stance detection. On Mawqif-v2 official public dev, raw Text-PRAS is positive across IG, GradShap, and LIME, but within-example shuffled baselines are consistently higher. This demonstrates that raw positive attribution-mass scores can overstate explanation alignment unless controlled against Atlas/token coverage. The negative baseline-adjusted finding is a meaningful audit result: it shows that the framework can distinguish raw positivity from evidence alignment above the within-example shuffled baseline.</p>
+<h2>References</h2>
+<ul>
+<li>Mohammad et al. (2016). SemEval-2016 Task 6: Detecting Stance in Tweets. https://aclanthology.org/S16-1003/</li>
+<li>Mawqif dataset (KFUPM). MAWQIF: A Multi-label Arabic Dataset for Target-specific Stance Detection. https://pure.kfupm.edu.sa/en/publications/mawqif-a-multi-label-arabic-dataset-for-target-specific-stance-de/</li>
+<li>StanceEval (2026). StanceEval shared task and MawqifV2 data page. https://github.com/StanceEval/stanceeval.github.io/tree/main/MawqifV2</li>
+<li>Baly et al. / AraStance (2021). AraStance: A Multi-Country and Multi-Domain Dataset of Arabic Stance Detection for Fact Checking. https://arxiv.org/abs/2104.13559</li>
+<li>Stance Prediction and Claim Verification (2020). Stance Prediction and Claim Verification: An Arabic Perspective. https://arxiv.org/abs/2005.10410</li>
+<li>Baly et al. (2018). Integrating Stance Detection and Fact Checking in a Unified Corpus. https://aclanthology.org/N18-1074/</li>
+<li>Alturayeif et al. (2022). ArSentD-LEV: A Multi-Topic Corpus for Target-based Sentiment Analysis in Arabic Levantine Tweets. https://arxiv.org/abs/1906.01830</li>
+<li>Abdul-Mageed et al. (2021). ARBERT &amp; MARBERT: Deep Bidirectional Transformers for Arabic. https://aclanthology.org/2021.acl-long.551/</li>
+<li>Antoun et al. (2020). AraBERT: Transformer-based Model for Arabic Language Understanding. https://arxiv.org/abs/2003.00104</li>
+<li>Abdelali et al. (2021). Pre-Training BERT on Arabic Tweets: Practical Considerations (QARiB). https://arxiv.org/abs/2102.10684</li>
+<li>AraELECTRA (2020). AraELECTRA: Pre-Training Text Discriminators for Arabic Language Understanding. https://arxiv.org/abs/2012.15516</li>
+<li>Zaidan et al. (2007). Using Annotator Rationales to Improve Machine Learning for Text Categorization. https://aclanthology.org/N07-1033/</li>
+<li>Lei et al. (2016). Rationalizing Neural Predictions. https://aclanthology.org/D16-1011/</li>
+<li>DeYoung et al. (2020). ERASER: A Benchmark to Evaluate Rationalized NLP Models. https://aclanthology.org/2020.acl-main.408/</li>
+<li>Camburu et al. (2018). e-SNLI: Natural Language Inference with Natural Language Explanations. https://arxiv.org/abs/1812.01193</li>
+<li>Jain et al. (2020). Learning to Faithfully Rationalize by Construction. https://aclanthology.org/2020.acl-main.409/</li>
+<li>Ribeiro et al. (2016). Why Should I Trust You?: Explaining the Predictions of Any Classifier. https://dl.acm.org/doi/10.1145/2939672.2939778</li>
+<li>Sundararajan et al. (2017). Axiomatic Attribution for Deep Networks. https://proceedings.mlr.press/v70/sundararajan17a.html</li>
+<li>Lundberg and Lee (2017). A Unified Approach to Interpreting Model Predictions. https://papers.nips.cc/paper/7062-a-unified-approach-to-interpreting-model-predictions</li>
+<li>Captum (2024). Captum GradientShap documentation. https://captum.ai/api/gradient_shap.html</li>
+<li>Jain and Wallace (2019). Attention is not Explanation. https://aclanthology.org/N19-1357/</li>
+<li>Wiegreffe and Pinter (2019). Attention is not not Explanation. https://aclanthology.org/D19-1002/</li>
+<li>Jacovi and Goldberg (2020). Towards Faithfully Interpretable NLP Systems. https://aclanthology.org/2020.acl-main.386/</li>
+<li>Adebayo et al. (2018). Sanity Checks for Saliency Maps. https://arxiv.org/abs/1810.03292</li>
+<li>Hooker et al. (2019). A Benchmark for Interpretability Methods in Deep Neural Networks. https://arxiv.org/abs/1806.10758</li>
+<li>Gururangan et al. (2018). Annotation Artifacts in Natural Language Inference Data. https://aclanthology.org/N18-2017/</li>
+<li>Poliak et al. (2018). Hypothesis Only Baselines in Natural Language Inference. https://aclanthology.org/S18-2023/</li>
+<li>McCoy et al. (2019). Right for the Wrong Reasons: Diagnosing Syntactic Heuristics in Natural Language Inference. https://aclanthology.org/P19-1334/</li>
+<li>Kaushik et al. (2020). Learning the Difference that Makes a Difference with Counterfactually-Augmented Data. https://arxiv.org/abs/1909.12434</li>
+<li>Geirhos et al. (2020). Shortcut learning in deep neural networks. https://www.nature.com/articles/s42256-020-00257-z</li>
+<li>Hase and Bansal (2020). Evaluating Explainable AI: Which Algorithmic Explanations Help Users Predict Model Behavior?. https://aclanthology.org/2020.acl-main.491/</li>
+<li>Pruthi et al. (2020). Learning to Deceive with Attention-Based Explanations. https://aclanthology.org/2020.acl-main.432/</li>
+</ul>
+</body></html>