{ "paper_id": "2022", "header": { "generated_with": "S2ORC 1.0.0", "date_generated": "2023-01-19T01:11:35.471170Z" }, "title": "UNIREX: A Unified Learning Framework for Language Model Rationale Extraction", "authors": [ { "first": "Aaron", "middle": [], "last": "Chan", "suffix": "", "affiliation": { "laboratory": "", "institution": "University of Southern California", "location": {} }, "email": "chanaaro@usc.edu" }, { "first": "Maziar", "middle": [], "last": "Sanjabi", "suffix": "", "affiliation": {}, "email": "maziars@fb.com" }, { "first": "Lambert", "middle": [], "last": "Mathias", "suffix": "", "affiliation": {}, "email": "mathiasl@fb.com" }, { "first": "Liang", "middle": [], "last": "Tan", "suffix": "", "affiliation": {}, "email": "liangtan@fb.com" }, { "first": "Shaoliang", "middle": [], "last": "Nie", "suffix": "", "affiliation": {}, "email": "snie@fb.com" }, { "first": "Xiaochang", "middle": [], "last": "Peng", "suffix": "", "affiliation": {}, "email": "xiaochang@fb.com" }, { "first": "Xiang", "middle": [], "last": "Ren", "suffix": "", "affiliation": { "laboratory": "", "institution": "University of Southern California", "location": {} }, "email": "xiangren@usc.edu" }, { "first": "Hamed", "middle": [], "last": "Firooz", "suffix": "", "affiliation": {}, "email": "mhfirooz@fb.com" } ], "year": "", "venue": null, "identifiers": {}, "abstract": "An extractive rationale explains a language model's (LM's) prediction on a given task instance by highlighting the text inputs that most influenced the prediction. Ideally, rationale extraction should be faithful (reflective of LM's actual behavior) and plausible (convincing to humans), without compromising the LM's (i.e., task model's) task performance. Although attribution algorithms and select-predict pipelines are commonly used in rationale extraction, they both rely on certain heuristics that hinder them from satisfying all three desiderata. In light of this, we propose UNIREX, a flexible learning framework which generalizes rationale extractor optimization as follows: (1) specify architecture for a learned rationale extractor; (2) select explainability objectives (i.e., faithfulness and plausibility criteria); and (3) jointly train the task model and rationale extractor on the task using selected objectives. UNIREX enables replacing prior works' heuristic design choices with a generic learned rationale extractor in (1) and optimizing it for all three desiderata in (2)-(3). To facilitate comparison between methods w.r.t. multiple desiderata, we introduce the Normalized Relative Gain (NRG) metric. Across five English text classification datasets, our best UNIREX configuration outperforms the strongest baselines by an average of 32.9% NRG. Plus, we find that UNIREXtrained rationale extractors' faithfulness can even generalize to unseen datasets and tasks.", "pdf_parse": { "paper_id": "2022", "_pdf_hash": "", "abstract": [ { "text": "An extractive rationale explains a language model's (LM's) prediction on a given task instance by highlighting the text inputs that most influenced the prediction. Ideally, rationale extraction should be faithful (reflective of LM's actual behavior) and plausible (convincing to humans), without compromising the LM's (i.e., task model's) task performance. Although attribution algorithms and select-predict pipelines are commonly used in rationale extraction, they both rely on certain heuristics that hinder them from satisfying all three desiderata. In light of this, we propose UNIREX, a flexible learning framework which generalizes rationale extractor optimization as follows: (1) specify architecture for a learned rationale extractor; (2) select explainability objectives (i.e., faithfulness and plausibility criteria); and (3) jointly train the task model and rationale extractor on the task using selected objectives. UNIREX enables replacing prior works' heuristic design choices with a generic learned rationale extractor in (1) and optimizing it for all three desiderata in (2)-(3). To facilitate comparison between methods w.r.t. multiple desiderata, we introduce the Normalized Relative Gain (NRG) metric. Across five English text classification datasets, our best UNIREX configuration outperforms the strongest baselines by an average of 32.9% NRG. Plus, we find that UNIREXtrained rationale extractors' faithfulness can even generalize to unseen datasets and tasks.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Abstract", "sec_num": null } ], "body_text": [ { "text": "Large neural language models (LMs) have yielded state-of-the-art performance on various natural language processing (NLP) tasks (Devlin et al., 2018; Liu et al., 2019) . However, LMs' complex reasoning processes are notoriously opaque (Rudin, 2019) , posing concerns about the societal implications of using LMs for high-stakes decision-making * Work done while AC was a research intern at Meta AI. (Bender et al., 2021) . Thus, explaining LMs' behavior is crucial for promoting trust, ethics, and safety in NLP systems (Doshi-Velez and Kim, 2017; Lipton, 2018) . Given a LM's (i.e., task model's) predicted label on a text classification instance, an extractive rationale is a type of explanation that highlights the tokens that most influenced the model to predict that label (Luo et al., 2021) . Ideally, rationale extraction should be faithful (Ismail et al., 2021; Jain et al., 2020) and plausible (DeYoung et al., 2019) , without hurting the LM's task performance (DeYoung et al., 2019) ( Fig. 1) .", "cite_spans": [ { "start": 128, "end": 149, "text": "(Devlin et al., 2018;", "ref_id": "BIBREF9" }, { "start": 150, "end": 167, "text": "Liu et al., 2019)", "ref_id": "BIBREF27" }, { "start": 235, "end": 248, "text": "(Rudin, 2019)", "ref_id": "BIBREF38" }, { "start": 399, "end": 420, "text": "(Bender et al., 2021)", "ref_id": "BIBREF2" }, { "start": 520, "end": 547, "text": "(Doshi-Velez and Kim, 2017;", "ref_id": "BIBREF11" }, { "start": 548, "end": 561, "text": "Lipton, 2018)", "ref_id": "BIBREF26" }, { "start": 778, "end": 796, "text": "(Luo et al., 2021)", "ref_id": "BIBREF29" }, { "start": 848, "end": 869, "text": "(Ismail et al., 2021;", "ref_id": "BIBREF16" }, { "start": 870, "end": 888, "text": "Jain et al., 2020)", "ref_id": "BIBREF18" }, { "start": 903, "end": 925, "text": "(DeYoung et al., 2019)", "ref_id": "BIBREF10" } ], "ref_spans": [ { "start": 995, "end": 1002, "text": "Fig. 1)", "ref_id": "FIGREF0" } ], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "Configuring the rationale extractor and its training can greatly impact these desiderata, yet prior works have commonly adopted two suboptimal heuristics. First, many works rely in some way on attribution algorithms (AAs), which extract rationales via handcrafted functions (Sundararajan et al., 2017; Ismail et al., 2021; Situ et al., 2021) . AAs cannot be directly trained and tend to be computeintensive (Bastings and Filippova, 2020) . Also, AAs can be a bottleneck for plausibility, as producing human-like rationales is a complex objec-tive requiring high capacity rationale extractors (Narang et al., 2020; DeYoung et al., 2019) . Second, many works use a specialized select-predict pipeline (SPP), where a predictor module is trained to solve the task using only tokens chosen by a selector module (Jain et al., 2020; Yu et al., 2021; . Instead of faithfulness optimization, SPPs heuristically aim for \"faithfulness by construction\" by treating the selected tokens as a rationale for the predictor's output (which depends only on those tokens). Still, SPPs typically have worse task performance than vanilla LMs since SPPs hide the full input from the predictor.", "cite_spans": [ { "start": 274, "end": 301, "text": "(Sundararajan et al., 2017;", "ref_id": "BIBREF45" }, { "start": 302, "end": 322, "text": "Ismail et al., 2021;", "ref_id": "BIBREF16" }, { "start": 323, "end": 341, "text": "Situ et al., 2021)", "ref_id": "BIBREF42" }, { "start": 407, "end": 437, "text": "(Bastings and Filippova, 2020)", "ref_id": "BIBREF1" }, { "start": 592, "end": 613, "text": "(Narang et al., 2020;", "ref_id": null }, { "start": 614, "end": 635, "text": "DeYoung et al., 2019)", "ref_id": "BIBREF10" }, { "start": 806, "end": 825, "text": "(Jain et al., 2020;", "ref_id": "BIBREF18" }, { "start": 826, "end": 842, "text": "Yu et al., 2021;", "ref_id": "BIBREF52" } ], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "To tackle this challenge, we propose the UNIfied Learning Framework for Rationale EXtraction (UNIREX), which generalizes rationale extractor optimization as follows: (1) specify architecture for a learned rationale extractor; (2) select explainability objectives (i.e., faithfulness and plausibility criteria); and (3) jointly train the task model and rationale extractor on the task using selected objectives (Sec. 3). UNIREX enables replacing prior works' heuristic design choices in (1) with a generic learned rationale extractor and optimizing it for all three desiderata in (2)-(3).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "UNIREX provides significant flexibility in performing (1)-(3). For (1), any model architecture is applicable, but we study Transformer LM based rationale extractors in this work (Zaheer et al., 2020; DeYoung et al., 2019) . We focus on two architectures: (A) Dual LM, where task model and rationale extractor are separate and (B) Shared LM, where task model and rationale extractor share parameters. For (2), any faithfulness and plausibility criteria can be used. Following DeYoung et al. (2019), we focus on comprehensiveness and sufficiency as faithfulness criteria, while using similarity to gold rationales as plausibility criteria. For (3), trade-offs between the three desiderata can be easily managed during rationale extractor optimization by setting arbitrary loss weights for the faithfulness and plausibility objectives. Plus, though computing the faithfulness criteria involves discrete (nondifferentiable) token selection, using Shared LM can approximate end-to-end training and enable both task model and rationale extractor to be optimized w.r.t. all three desiderata (Sec. 3.3).", "cite_spans": [ { "start": 178, "end": 199, "text": "(Zaheer et al., 2020;", "ref_id": "BIBREF53" }, { "start": 200, "end": 221, "text": "DeYoung et al., 2019)", "ref_id": "BIBREF10" } ], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "To evaluate all three desiderata in aggregate, we introduce the Normalized Relative Gain (NRG) metric. Across five English text classification datasets -SST, Movies, CoS-E, MultiRC, and e-SNLI (Carton et al., 2020; DeYoung et al., 2019 )our best UNIREX configuration outperforms the strongest baselines by an average of 32.9% NRG (Sec. 4.2) , showing that UNIREX can optimize rationale extractors for all three desiderata. In addition, we verify our UNIREX design choices via extensive ablation studies (Sec. 4.3) . Furthermore, UNIREX-trained extractors have high generalization power, yielding high plausiblity with minimal gold rationale supervision (Sec. 4.4) and high faithfulness on unseen datasets and tasks (Sec. 4.5).", "cite_spans": [ { "start": 193, "end": 214, "text": "(Carton et al., 2020;", "ref_id": null }, { "start": 215, "end": 235, "text": "DeYoung et al., 2019", "ref_id": "BIBREF10" }, { "start": 330, "end": 340, "text": "(Sec. 4.2)", "ref_id": null }, { "start": 503, "end": 513, "text": "(Sec. 4.3)", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "Finally, our user study shows that humans judge UNIREX rationales as more plausible than rationales extracted using other methods (Sec. 4.6).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "Rationale Extraction Let F task = f task (f enc (\u2022)) be a task model for M -class text classification (Sec. A.1), where f enc is the text encoder and f task is the task output head. Typically, F task has a BERT-style architecture (Devlin et al., 2018) , in which f enc is a Transformer (Vaswani et al., 2017) while f task is a linear layer with softmax classifier. Let", "cite_spans": [ { "start": 230, "end": 251, "text": "(Devlin et al., 2018)", "ref_id": "BIBREF9" }, { "start": 286, "end": 308, "text": "(Vaswani et al., 2017)", "ref_id": "BIBREF48" } ], "ref_spans": [], "eq_spans": [], "section": "Problem Formulation", "sec_num": "2" }, { "text": "x i = [x t i ]", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Problem Formulation", "sec_num": "2" }, { "text": "n t=1 be the n-token input sequence (e.g., a sentence) for task instance i, and F task (x i ) \u2208 R M be the logit vector for the output of the task model. Let\u0177 i = arg max j F task (x i ) j be the class predicted by F task . Given F task , x i , and\u0177 i , the goal of rationale extraction is to output vector", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Problem Formulation", "sec_num": "2" }, { "text": "s i = [s t i ] n t=1 \u2208 R n , such that each s t i \u2208 R", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Problem Formulation", "sec_num": "2" }, { "text": "is an importance score indicating how much token x t i influenced F task to predict class\u0177 i . Let F ext be a rationale extractor, such that", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Problem Formulation", "sec_num": "2" }, { "text": "s i = F ext (F task , x i ,\u0177 i ).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Problem Formulation", "sec_num": "2" }, { "text": "F ext can be a learned or heuristic function. In practice, the final rationale is often obtained by binarizing s i as r i \u2208 {0, 1} n , via the top-k% strategy: r t i = 1 if s t i is one of the top-k% scores in s i ; otherwise, r t i = 0 (DeYoung et al., 2019; Jain et al., 2020; Pruthi et al., 2020; Chan et al., 2021) . For top-k%, let r (k) i be the \"important\" (i.e., ones) tokens in r i , when using 0 \u2264 k \u2264 100.", "cite_spans": [ { "start": 237, "end": 259, "text": "(DeYoung et al., 2019;", "ref_id": "BIBREF10" }, { "start": 260, "end": 278, "text": "Jain et al., 2020;", "ref_id": "BIBREF18" }, { "start": 279, "end": 299, "text": "Pruthi et al., 2020;", "ref_id": "BIBREF36" }, { "start": 300, "end": 318, "text": "Chan et al., 2021)", "ref_id": "BIBREF6" } ], "ref_spans": [], "eq_spans": [], "section": "Problem Formulation", "sec_num": "2" }, { "text": "Faithfulness means how well a rationale reflects F task 's true reasoning process for predicting\u0177 i (Jacovi and Goldberg, 2020) . Hence, faithfulness metrics measure how much the r (k) i tokens impact p\u0177 i (x i ), which denotes F task 's confidence probability for\u0177 i when using x i as input (DeYoung et al., 2019; Shrikumar et al., 2017; Hooker et al., 2018; Pruthi et al., 2020) . Recently, comprehensiveness and sufficiency have emerged as popular faithfulness metrics (DeYoung et al., 2019) . Comprehensiveness (comp) measures the change in p\u0177 i when r (k) i is removed from the input:", "cite_spans": [ { "start": 100, "end": 127, "text": "(Jacovi and Goldberg, 2020)", "ref_id": "BIBREF17" }, { "start": 292, "end": 314, "text": "(DeYoung et al., 2019;", "ref_id": "BIBREF10" }, { "start": 315, "end": 338, "text": "Shrikumar et al., 2017;", "ref_id": "BIBREF40" }, { "start": 339, "end": 359, "text": "Hooker et al., 2018;", "ref_id": "BIBREF15" }, { "start": 360, "end": 380, "text": "Pruthi et al., 2020)", "ref_id": "BIBREF36" }, { "start": 472, "end": 494, "text": "(DeYoung et al., 2019)", "ref_id": "BIBREF10" } ], "ref_spans": [], "eq_spans": [], "section": "Problem Formulation", "sec_num": "2" }, { "text": "comp = p\u0177 i (x i ) \u2212 p\u0177 i (x i \\r (k) i ). Sufficiency (suff) measures the change in p\u0177 i when only r (k) i is kept in the input: suff = p\u0177 i (x i ) \u2212 p\u0177 i (r (k) i ).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Problem Formulation", "sec_num": "2" }, { "text": "High faithfulness is signaled by high comp and low suff.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Problem Formulation", "sec_num": "2" }, { "text": "Plausibility means how convincing a rationale is to humans (Jacovi and Goldberg, 2020) . This can be measured by automatically computing the similarity between F ext 's rationales (either s i or r i ) and human-annotated gold rationales (DeYoung et al., 2019) , or by asking human annotators to rate whether F ext 's rationales make sense for predicting\u0177 i (Strout et al., 2019; Doshi-Velez and Kim, 2017) . Typically, a gold rationale is a binary vector r * i \u2208 {0, 1} n , where ones/zeros indicate important/unimportant tokens (Lei et al., 2016) .", "cite_spans": [ { "start": 59, "end": 86, "text": "(Jacovi and Goldberg, 2020)", "ref_id": "BIBREF17" }, { "start": 237, "end": 259, "text": "(DeYoung et al., 2019)", "ref_id": "BIBREF10" }, { "start": 357, "end": 378, "text": "(Strout et al., 2019;", "ref_id": "BIBREF44" }, { "start": 379, "end": 405, "text": "Doshi-Velez and Kim, 2017)", "ref_id": "BIBREF11" }, { "start": 529, "end": 547, "text": "(Lei et al., 2016)", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "Problem Formulation", "sec_num": "2" }, { "text": "Task Performance, w.r.t. rationale extraction, concerns how much F task 's task performance (on test set) drops when F task is trained with explainability objectives (i.e., faithfulness, plausibility) for F ext . As long as F task is trained with non-task losses, F task 's task performance can be affected.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Problem Formulation", "sec_num": "2" }, { "text": "Given task model F task , UNIREX generalizes rationale extractor optimization as follows: (1) choose architecture for a learned rationale extractor F ext ; (2) select explainability objectives (i.e., faithfulness loss L faith and plausibility loss L plaus ); and (3) jointly train F task and F ext using L task (task loss), L faith , and L plaus . UNIREX training consists of two backpropagation paths (Fig. 2) . The first path is used to update F task w.r.t. L task and L faith . Whereas L task is computed w.r.t. the task target y i , L faith is computed only using the task input x i and the top-k% important tokens r (k) i (obtained via F ext ), based on some combination of comp and suff (Sec. 2). The second path is used to update F ext w.r.t. L plaus , which encourages importance scores s i to approximate gold rationale r * i . Thus, UNIREX frames rationale extraction as the following optimization problem:", "cite_spans": [], "ref_spans": [ { "start": 402, "end": 410, "text": "(Fig. 2)", "ref_id": "FIGREF1" } ], "eq_spans": [], "section": "UNIREX", "sec_num": "3" }, { "text": "EQUATION", "cite_spans": [], "ref_spans": [], "eq_spans": [ { "start": 0, "end": 8, "text": "EQUATION", "ref_id": "EQREF", "raw_str": "min F task , Fext L task (x i , y i ; F task ) + \u03b1 f L faith (x i , r (k) i ; F task ) + \u03b1 p L plaus (x i , r * i ; F ext ),", "eq_num": "(1)" } ], "section": "UNIREX", "sec_num": "3" }, { "text": "where \u03b1 f and \u03b1 p are loss weights. If F task and F ext share parameters, then the shared parameters will be optimized w.r.t. all losses. During inference, for task input x i , we first use F task to predict y i , then use F ext to output a rationale r i for F task 's prediction\u0177 i . Below, we discuss options for the rationale extractor and explainability objectives.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "UNIREX", "sec_num": "3" }, { "text": "In UNIREX, F ext is a learned function by default. Learned F ext can be any model that transforms x t i into s t i . Given their success in NLP explainability (DeYoung et al., 2019), we focus on pre-trained Transformer LMs and highlight two architectures: Dual LM (DLM) and Shared LM (SLM) (Fig. 3) . For DLM, F task and F ext are two separate Transformer LMs. DLM provides more dedicated capacity for F ext , which can help F ext output plausible rationales. For SLM, F task and F ext are two Transformer LMs sharing encoder f enc , while F ext has its own output head f ext . SLM leverages multitask learning between F task and F ext , which can improve faithfulness since F ext gets more information about F task 's reasoning process. Unlike heuristic F ext (Sec. A.2), learned F ext can be optimized for faithfulness/plausibility, but cannot be used out of the box without training. Learned F ext is preferred if: (A) optimizing for both faithfulness and plausibility, and (B) gold rationales are available for plausibility optimization (Sec. A.3).", "cite_spans": [], "ref_spans": [ { "start": 290, "end": 298, "text": "(Fig. 3)", "ref_id": "FIGREF2" } ], "eq_spans": [], "section": "Rationale Extractor", "sec_num": "3.1" }, { "text": "After selecting F ext , we specify the explainability objectives, which can be any combination of faithfulness and plausibility criteria. In prior approaches (e.g., AA, SPPs), the rationale extractor is not optimized for both faithfulness and plausibility, but UNIREX makes this possible. For any choice of learned F ext , UNIREX lets us easily \"plug and play\" different criteria and loss weights, based on our needs and domain knowledge, to find those that best balance the rationale extraction desiderata.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Explainability Objectives", "sec_num": "3.2" }, { "text": "Faithfulness Evaluating rationale faithfulness is still an open problem with many existing metrics, and UNIREX is not tailored for any specific metric. Still, given the prevalence of comp/suff (Sec. 2), we focus on comp/suff based objectives.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Explainability Objectives", "sec_num": "3.2" }, { "text": "Recall that comp measures the importance of tokens in r (k) i as how p\u0177 i (x i ), F task 's predicted probability for class\u0177 i , changes when those tokens are removed from x i . Intuitively, we want p\u0177 i (x i ) to be higher than p\u0177 i (x i \\r (k) i ), so higher comp is better. Since comp is defined for a single class' probability rather than the label distribution, we can define the comp loss L comp via cross-entropy loss L CE , as in the following difference criterion for L comp :", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Explainability Objectives", "sec_num": "3.2" }, { "text": "EQUATION", "cite_spans": [], "ref_spans": [], "eq_spans": [ { "start": 0, "end": 8, "text": "EQUATION", "ref_id": "EQREF", "raw_str": "Lcomp-diff = LCE(Ftask(xi), yi) \u2212 LCE(Ftask(xi\\r (k) i ), yi)) (2) L CE (F task (xi), yi) = \u2212yi log(F task (xi))", "eq_num": "(3)" } ], "section": "Explainability Objectives", "sec_num": "3.2" }, { "text": "For training stability, we compute comp loss for target class y i here instead of F task 's predicted class\u0177 i , since\u0177 i is a moving target during training. Using L comp-diff , it is possible for", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Explainability Objectives", "sec_num": "3.2" }, { "text": "L CE (F task (x i \\r (k) i ), y i )) to become much larger than L CE (F task (x i ), y i )", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Explainability Objectives", "sec_num": "3.2" }, { "text": ", leading to arbitrarily negative losses. To avoid this, we can add margin m c to the loss function, giving the margin criterion:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Explainability Objectives", "sec_num": "3.2" }, { "text": "Lcomp-margin = max(\u2212mc, LCE(Ftask(xi), yi) \u2212 LCE(Ftask(xi\\r (k) i ), yi)) + mc (4)", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Explainability Objectives", "sec_num": "3.2" }, { "text": "Recall that suff measures the importance of tokens in r (k) i as how p\u0177 i (x i ), F task 's predicted probability for class\u0177 i , changes when they are the only tokens kept in x i . Based on suff's definition, we want p\u0177 i (r (k) i ) to be higher than p\u0177 i (x i ), so lower suff is better. For suff loss L suff , we define the difference and margin criteria analogously with margin m s but the opposite sign (since lower suff is better):", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Explainability Objectives", "sec_num": "3.2" }, { "text": "EQUATION", "cite_spans": [], "ref_spans": [], "eq_spans": [ { "start": 0, "end": 8, "text": "EQUATION", "ref_id": "EQREF", "raw_str": "Lsuff-diff = LCE(Ftask(r (k) i ), yi) \u2212 LCE(Ftask(xi), yi) (5) Lsuff-margin = max(\u2212ms, LCE(Ftask(r (k) i ), yi) \u2212 LCE(Ftask(xi), yi)) + ms", "eq_num": "(6)" } ], "section": "Explainability Objectives", "sec_num": "3.2" }, { "text": "In our experiments, we find that the marginbased comp/suff criteria are effective (Sec. 4.3), though others (e.g., KL Div, MAE) can be used too (Sec. A.4.1). Note that r (k) i is computed via top-k% thresholding (Sec. 2), so we also need to specify a set K of threshold values. We separately compute the comp/suff losses for each k \u2208 K, then obtain the final comp/suff losses by averaging over all k values via area-over-precision-curve (AOPC) (DeYoung et al., 2019). To reflect this, we denote the comp and suff losses as L comp,K and L suff,K , respectively. Let", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Explainability Objectives", "sec_num": "3.2" }, { "text": "\u03b1 f L faith = \u03b1 c L comp,K + \u03b1 s L suff,K ,", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Explainability Objectives", "sec_num": "3.2" }, { "text": "where \u03b1 c and \u03b1 s are loss weights.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Explainability Objectives", "sec_num": "3.2" }, { "text": "Plausibility Plausibility is defined as how convincing a rationale is to humans (Jacovi and Goldberg, 2020), i.e., whether humans would agree the rationale supports the model's prediction. While optimizing for plausibility should ideally involve human-in-the-loop feedback, this is prohibitive. Instead, many works consider gold rationales as a cheaper form of plausibility annotation (DeYoung et al., 2019; Narang et al., 2020; Jain et al., 2020) . Thus, if gold rationale supervision is available, then we can optimize for plausibility. With gold rationale r * i for input x i , plausibility optimization entails training F ext to predict binary importance label r * ,t i for each token x t i . This is essentially token classification, so one natural choice for L plaus is the token-level binary cross-entropy (BCE) criterion:", "cite_spans": [ { "start": 385, "end": 407, "text": "(DeYoung et al., 2019;", "ref_id": "BIBREF10" }, { "start": 408, "end": 428, "text": "Narang et al., 2020;", "ref_id": null }, { "start": 429, "end": 447, "text": "Jain et al., 2020)", "ref_id": "BIBREF18" } ], "ref_spans": [], "eq_spans": [], "section": "Explainability Objectives", "sec_num": "3.2" }, { "text": "EQUATION", "cite_spans": [], "ref_spans": [], "eq_spans": [ { "start": 0, "end": 8, "text": "EQUATION", "ref_id": "EQREF", "raw_str": "Lplaus-BCE = \u2212 t r * ,t i log(Fext(x t i ))", "eq_num": "(7)" } ], "section": "Explainability Objectives", "sec_num": "3.2" }, { "text": "Besides BCE loss, we can also consider other criteria like sequence-level KL divergence and L1 loss. See Sec. A.4.2 for discussion of these and other plausibility criteria.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Explainability Objectives", "sec_num": "3.2" }, { "text": "After setting F ext , L faith , and L plaus , we can move on to training F task and F ext . Since top-k% rationale binarization (Sec. 3.2) is not differentiable, by default, we cannot backpropagate L faith through all of F ext 's parameters. Thus, F task is trained via L task and L faith , while F ext is only trained via L plaus . This means F ext 's rationales r i are indirectly optimized for faithfulness by regularizing F task such that its behavior aligns with r i . The exception is if we are using the SLM variant, where encoder f enc is shared by F task and F ext . In this case, f enc is optimized w.r.t. all losses, f task is optimized w.r.t. L task and L faith , and f ext is optimized w.r.t. L plaus . SLM is a simple way to approximate end-to-end training of F task and F ext . In contrast, past SPPs have used more complex methods like reinforcement learning (Lei et al., 2016) and the reparameterization trick (Bastings et al., 2019) , whose training instability can hurt task performance (Jain et al., 2020) . Now, we summarize the full learning objective. Given that cross-entropy loss L task = L CE (F task (x i ), y i ) is used to train F task to predict y i , the full learning objective is:", "cite_spans": [ { "start": 875, "end": 893, "text": "(Lei et al., 2016)", "ref_id": null }, { "start": 927, "end": 950, "text": "(Bastings et al., 2019)", "ref_id": "BIBREF0" }, { "start": 1006, "end": 1025, "text": "(Jain et al., 2020)", "ref_id": "BIBREF18" } ], "ref_spans": [], "eq_spans": [], "section": "Training and Inference", "sec_num": "3.3" }, { "text": "EQUATION", "cite_spans": [], "ref_spans": [], "eq_spans": [ { "start": 0, "end": 8, "text": "EQUATION", "ref_id": "EQREF", "raw_str": "L = Ltask + \u03b1 f Lfaith + \u03b1pLplaus = Ltask + \u03b1cLcomp,K + \u03b1sLsuff,K + \u03b1pLplaus.", "eq_num": "(8)" } ], "section": "Training and Inference", "sec_num": "3.3" }, { "text": "During inference, we use F task to predict y i , then use F ext to output r i for F task 's predicted label\u0177 i .", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Training and Inference", "sec_num": "3.3" }, { "text": "We present empirical results demonstrating UNIREX's effectiveness in managing trade-offs between faithfulness, plausibility, and task performance during rationale extractor optimization. First, our main experiments compare methods w.r.t. faithfulness, plausibility, and task performance (Sec. 4.2). Second, we perform various ablation studies to verify our design choices for UNIREX (Sec. 4.3). Third, we present experiments highlighting UNIREX's generalization ability, both in terms of limited gold rationale supervision (Sec. 4.4) and zero-shot transfer (Sec. 4.5). Fourth, we conduct a user study to further evaluate UNIREX rationales' plausibility, relative to those generated by other methods (Sec. 4.6). See Sec. A.5 for implementation details (LM architecture, AA settings, training).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Experiments", "sec_num": "4" }, { "text": "Datasets We primarily use SST (Socher et al., 2013; Carton et al., 2020) , Movies (Zaidan and Eisner, 2008) , CoS-E (Rajani et al., 2019), Mul-tiRC (Khashabi et al., 2018) , and e-SNLI (Camburu et al., 2018) , all of which have gold rationale annotations. The latter four datasets were taken from the ERASER benchmark (DeYoung et al., 2019).", "cite_spans": [ { "start": 30, "end": 51, "text": "(Socher et al., 2013;", "ref_id": "BIBREF43" }, { "start": 52, "end": 72, "text": "Carton et al., 2020)", "ref_id": null }, { "start": 82, "end": 107, "text": "(Zaidan and Eisner, 2008)", "ref_id": "BIBREF54" }, { "start": 148, "end": 171, "text": "(Khashabi et al., 2018)", "ref_id": "BIBREF20" }, { "start": 185, "end": 207, "text": "(Camburu et al., 2018)", "ref_id": "BIBREF4" } ], "ref_spans": [], "eq_spans": [], "section": "Experiment Setup", "sec_num": "4.1" }, { "text": "We use the metrics from the ERASER explainability benchmark (DeYoung et al., 2019) . For faithfulness, we use comprehensiveness (Comp) and sufficiency (Suff), for k = [1, 5, 10, 20, 50] (DeYoung et al., 2019). For plausibility, we use area under precision-recall curve (AUPRC) and token F1 (TF1) to measure similarity to gold rationales (DeYoung et al., 2019; Narang et al., 2020) . For task performance, we follow (DeYoung et al., 2019) and (Carton et al., 2020) in using accuracy (SST, CoS-E) and macro F1 (Movies, MultiRC, e-SNLI).", "cite_spans": [ { "start": 60, "end": 82, "text": "(DeYoung et al., 2019)", "ref_id": "BIBREF10" }, { "start": 337, "end": 359, "text": "(DeYoung et al., 2019;", "ref_id": "BIBREF10" }, { "start": 360, "end": 380, "text": "Narang et al., 2020)", "ref_id": null }, { "start": 442, "end": 463, "text": "(Carton et al., 2020)", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "Metrics", "sec_num": null }, { "text": "To aggregately evaluate multiple desiderata, we introduce the Normalized Relative Gain (NRG) metric, which is based on the ARG metric from Ye et al. (2021) . NRG normalizes raw metrics (e.g., F1, sufficiency) to scores between 0 and 1 (higher is better). Given a set of raw metric scores Z = {z 1 , z 2 , ...} (each from a different method), NRG(z i ) captures z i 's value relative to min(Z) and max(Z). If higher values are better for the given metric (e.g., F1), then we have:", "cite_spans": [ { "start": 139, "end": 155, "text": "Ye et al. (2021)", "ref_id": "BIBREF50" } ], "ref_spans": [], "eq_spans": [], "section": "Metrics", "sec_num": null }, { "text": "NRG(z i ) = z i \u2212min(Z) max(Z)\u2212min(Z)", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Metrics", "sec_num": null }, { "text": ". If lower values are better (e.g., sufficiency), then we have:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Metrics", "sec_num": null }, { "text": "NRG(z i ) = max(Z)\u2212z i max(Z)\u2212min(Z)", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Metrics", "sec_num": null }, { "text": ". After computing NRG for multiple raw metrics, we can aggregate them w.r.t. desiderata via averaging. Let FNRG, PNRG, and TNRG be the NRG values for faithfulness, plausibility, and task performance, respectively. Finally, we compute the composite NRG as:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Metrics", "sec_num": null }, { "text": "CNRG = FNRG+PNRG+TNRG 3", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Metrics", "sec_num": null }, { "text": ". Results Reporting For all results, we report average over three seeds and the five k values. We Figure 4 : Composite NRG Comparison (w/o Plausibility Optimization). Composite NRG (CNRG) is the mean of the three desiderata NRG scores. For each dataset, we use CNRG to compare methods that do not optimize for plausibility. Figure 5 : Composite NRG Comparison (w/ Plausibility Optimization). Composite NRG (CNRG) is the mean of the three desiderata NRG scores. For each dataset, we use CNRG to compare methods that do optimize for plausibility. denote each UNIREX configuration with \"([rationale extractor]-[explainability objectives])\". F, P, and FP denote faithfulness, plausibility, and faith-fulness+plausibility, respectively.", "cite_spans": [], "ref_spans": [ { "start": 98, "end": 106, "text": "Figure 4", "ref_id": null }, { "start": 324, "end": 332, "text": "Figure 5", "ref_id": null } ], "eq_spans": [], "section": "Metrics", "sec_num": null }, { "text": "Baselines The first category is AAs, which are not trained: AA (Grad) (Simonyan et al., 2013) , AA (Input*Grad) (Denil et al., 2014) , AA (DeepLIFT) (Lundberg and Lee, 2017), AA (IG) (Sundararajan et al., 2017) . We also experiment with IG for L2E (Situ et al., 2021) , which distills knowledge from an AA to an LM. The second category is SPPs: FRESH (Jain et al., 2020) and A2R (Yu et al., 2021) . For FRESH, we use a strong variant where IG rationales are directly given to the predictor, rather than output by a trained selector. A2R aims to improve SPP task performance by regularizing the predictor with an attention-based predictor that uses the full input. In addition, we introduce FRESH+P and A2R+P, which augment FRESH and A2R, respectively, with plausibility optimization. The third category is AA-based regularization: SGT (Ismail et al., 2021) , which uses a sufficiency-based criterion to optimize for faithfulness. We also consider SGT+P, which augments SGT with plausibility optimization. Fig. 4 -6 display the main results. In Fig. 4 /5, we compare the CNRG for all methods and datasets, without/with gold rationales. In both plots, we see that UNIREX variants achieve the best CNRG across all datasets, indicating that they are effective in balancing the three desiderata. In particular, UNIREX (DLM-FP) and UNIREX (SLM- FP) have very high CNRG scores, both yielding more than 30% improvement over the strongest baselines. Fig. 6 compares methods w.r.t. desiderata NRG (i.e., FNRG, PNRG, TNRG). Here, the left/right plots show methods without/with gold rationales. Again, we see that UNIREX variants achieve a good NRG balance of faithfulness, plausibility, and task performance. Meanwhile, many baselines (e.g., AA (IG), A2R, SGT+P) do well on some desiderata but very poorly on others.", "cite_spans": [ { "start": 70, "end": 93, "text": "(Simonyan et al., 2013)", "ref_id": "BIBREF41" }, { "start": 112, "end": 132, "text": "(Denil et al., 2014)", "ref_id": "BIBREF8" }, { "start": 183, "end": 210, "text": "(Sundararajan et al., 2017)", "ref_id": "BIBREF45" }, { "start": 248, "end": 267, "text": "(Situ et al., 2021)", "ref_id": "BIBREF42" }, { "start": 351, "end": 370, "text": "(Jain et al., 2020)", "ref_id": "BIBREF18" }, { "start": 379, "end": 396, "text": "(Yu et al., 2021)", "ref_id": "BIBREF52" }, { "start": 835, "end": 856, "text": "(Ismail et al., 2021)", "ref_id": "BIBREF16" } ], "ref_spans": [ { "start": 1005, "end": 1011, "text": "Fig. 4", "ref_id": null }, { "start": 1044, "end": 1050, "text": "Fig. 4", "ref_id": null }, { "start": 1441, "end": 1447, "text": "Fig. 6", "ref_id": "FIGREF3" } ], "eq_spans": [], "section": "Metrics", "sec_num": null }, { "text": "We present five ablation studies to validate the effectiveness of our UNIREX design choices. The ablation results are displayed in Table 1 . In this table, each of the five sections shows results for a different ablation. Thus, all numbers within the same section and column are comparable.", "cite_spans": [], "ref_spans": [ { "start": 131, "end": 138, "text": "Table 1", "ref_id": "TABREF1" } ], "eq_spans": [], "section": "Ablation Studies", "sec_num": "4.3" }, { "text": "Extractor Type In the Ext Type (F) section, we compare four heuristic rationale extractors, using AA-F. Rand uses random importance scores, Gold directly uses the gold rationales, Inv uses the inverse of the gold rationales, and IG uses IG. All heuristics yield similar task performance, but IG dominates on all faithfulness metrics. This makes sense because IG is computed using F task 's inputs/parameters/outputs, while the others do not have this information. For plausibility, Gold is the best, Inv is the worst, and Rand and IG are about the same, as none of the heuristics are optimized for plausibility. In the Ext Type (FP) section, we compare four learned rationale extractors. By default, attribution algorithms' dimension scores are pooled into token scores via sum pooling. AA-FP (Sum) uses IG with sum pooling, while AA-FP (MLP) replaces the sum pooler with a MLP-based pooler to increase capacity for plausibility optimization. Task performance for all four methods is similar, AA-FP (Sum) dominates on faithfulness, and DLM-FP and SLM-FP dominate on plausibility. AA-FP (MLP) does not perform as well on faithfulness but slightly improves on plausibility compared to AA-FP (Sum).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Ablation Studies", "sec_num": "4.3" }, { "text": "Comp/Suff Losses The Comp/Suff Loss section compares different combinations of Comp and Suff losses, using SLM-FP. Note that SLM-FP (Comp+Suff) is equivalent to SLM-FP shown in other tables/sections. As expected, SLM-FP (Comp) does best on Comp, but SLM-FP (Comp+Suff) actually does best on Suff. Meanwhile, SLM-FP, (Suff) does second-best on Suff but is much worse on Comp. This shows that Comp and Suff are complementary for optimization.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Ablation Studies", "sec_num": "4.3" }, { "text": "Suff Criterion The Suff Criterion section compares different Suff criteria, using SLM-FP. SLM-FP (KLDiv) uses the KL divergence criterion, SLM-FP (MAE) uses the MAE criterion, and SLM-FP (Margin) uses the margin criterion. SLM-FP (Margin) is equivalent to SLM-FP in other ta- bles/sections. All criteria yield similar performance and plausibility, while Margin is slightly better on faithfulness.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Ablation Studies", "sec_num": "4.3" }, { "text": "SLM Extractor Head The SLM Ext Head section compares different extractor heads, using SLM-FP. Linear is the default choice and uses a linear layer. MLP-2048-2 uses a MLP with two 2048-dim hidden layers. MLP-4096-3 uses a MLP with three 4096-dim hidden layers. All three output head types yield similar performance, but decreasing head capacity yields better faithfulness, while increasing head capacity heads yields better plausibility. This trades off faithfulness and plausibility, although larger heads will be more compute-intensive.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Ablation Studies", "sec_num": "4.3" }, { "text": "UNIREX supports arbitrary amounts of gold rationale supervision and allows us to account for data efficiency. In Fig. 7 , we compare plausibility (in AUPRC) for \u03b3 = [0.5, 1, 5, 10, 20, 100] (i.e., % of train instances with gold rationales). We compare AA (IG) and four UNIREX variants (AA-F, AA-FP, DLM-FP, SLM-FP). AA (IG) and AA-F do not use gold rationales and thus have the same AUPRC for all \u03b3. Standard deviation is shown by the error bands. UNIREX (DLM-FP) and UNIREX (SLM-FP) dominate across all \u03b3 values, with AUPRC slowly decreasing as \u03b3 decreases. Even at \u03b3 = 0.5, they can still achieve high AUPRC. This suggests that UNIREX's gold rationale batching procedure (Sec. A.3) is effective for learning from minimal gold rationale supervision and demonstrates how UNIREX enables us to manage this trade-off. See Sec. A.6 for similar results on CoS-E. ", "cite_spans": [], "ref_spans": [ { "start": 113, "end": 119, "text": "Fig. 7", "ref_id": "FIGREF4" } ], "eq_spans": [], "section": "Gold Rationale Data Efficiency", "sec_num": "4.4" }, { "text": "In Table 2 , we investigate if F ext 's faithfulness, via UNIREX training on some source dataset, can generalize to unseen target datasets/tasks in a zero-shot setting (i.e., no fine-tuning on target datasets). Plausibility is not evaluated here, since these unseen datasets do not have gold rationales. As the source model, we compare various SST-trained models: AA (IG) and UNIREX (AA-F, DLM-FP). First, we evaluate on unseen datasets for a seen task (sentiment analysis (SA)): Yelp (Zhang et al., 2015) and Amazon (McAuley and Leskovec, 2013) . Second, we evaluate on unseen datasets for unseen tasks: Stormfront (hate speech detection (HSD), binary F1) (de Gibert et al., 2018), OffenseEval (offensive speech detection (OSD), macro F1) (Zampieri et al., 2019) , and SemEval2018 (irony detection (ID), binary F1) (Van Hee et al., 2018). We want to show that, even if F task yields poor task performance on unseen datasets, F ext 's rationales can still be faithful. As expected, all methods achieve much lower task performance in the third setting than in the first two settings. However, faithfulness does not appear to be strongly correlated with task performance, as unseen tasks' comp/suff scores are similar to seen tasks'. Across all datasets, DLM-FP has the best faithfulness and is the only method whose comp is always higher than suff. AA-F is not as consistently strong as DLM-FP, but almost always beats AA (IG) on comp and suff. Meanwhile, AA (IG) has the worst comp and suff overall. Ultimately, these results suggest that UNIREX-trained models' faithfulness (i.e., alignment between F task 's and F ext 's outputs) is a dataset/task agnostic property (i.e., can generalize across datasets/tasks), further establishing UNIREX's utility in low-resource settings. ", "cite_spans": [ { "start": 485, "end": 505, "text": "(Zhang et al., 2015)", "ref_id": "BIBREF56" }, { "start": 510, "end": 545, "text": "Amazon (McAuley and Leskovec, 2013)", "ref_id": null }, { "start": 740, "end": 763, "text": "(Zampieri et al., 2019)", "ref_id": "BIBREF55" } ], "ref_spans": [ { "start": 3, "end": 10, "text": "Table 2", "ref_id": "TABREF3" } ], "eq_spans": [], "section": "Zero-Shot Faithfulness Transfer", "sec_num": "4.5" }, { "text": "Gold rationale based plausibility evaluation is noisy because gold rationales are for the target label, not a model's predicted label. Thus, we conduct two five-annotator user studies (Table 3) to get a better plausibility measurement. Given 50 random test instances from SST, we get the rationales for SGT+P, A2R+P, UNIREX (AA-FP), and UNIREX (DLM-FP), plus the gold rationales. For each instance, we threshold all rationales to have the same number of positive tokens as the gold rationale. The first user study is forward simulation (Hase and Bansal, 2020; Jain et al., 2020) . Here, the annotator is given an input and a rationale for some model's prediction, then asked what (binary) sentiment label the model most likely predicted. For forward simulation, we also consider a No Rationale baseline, where no tokens are highlighted. For No Rationale and Gold, the target label is the correct choice. Annotators are also asked to rate their confidence (4point Likert scale) in their answer to this question. The second user study involves giving a subjective rating of how plausible the rationale is (Hase and Bansal, 2020) . Here, the annotator is given the input, rationale, and model's predicted label, then asked to rate (5-point Likert scale) how aligned the rationale is with the prediction. In both forward simulation and subjective rating, we find that DLM-FP performs best among all non-oracle methods and even beats Gold on accuracy, further supporting that DLM-FP rationales are plausible. As expected, the fact that Gold does not achieve near-100% accuracy shows the discrepancy between evaluating plausibility based on the target label (i.e., gold rationale similarity) and F task 's predicted label (forward simulation). Meanwhile, SGT+P and AA-FP, which had lower AUPRC/TF1 in our automatic evaluation, also do worse in accuracy/alignment. Also, users found SGT+P and AA-FP rationales harder to understand, as shown by their lower confidence scores. Meanwhile, A2R+P had high AUPRC/TF1, but gets very low accuracy/alignment because A2R+P's predicted label often not the target label, leading to misalignment with its gold-like rationale. A2R+P is a great example of how automatic plausibility evaluation can be misleading. For the accuracy, confidence, and alignment questions, we achieved Fleiss' Kappa (Fleiss, 1971) inter-annotator agreement scores of 0.2456 (fair), 0.1282 (slight), and, 0.1561 (slight), respectively. This lack of agreement shows the difficulty of measuring plausibility.", "cite_spans": [ { "start": 536, "end": 559, "text": "(Hase and Bansal, 2020;", "ref_id": "BIBREF14" }, { "start": 560, "end": 578, "text": "Jain et al., 2020)", "ref_id": "BIBREF18" }, { "start": 1103, "end": 1126, "text": "(Hase and Bansal, 2020)", "ref_id": "BIBREF14" }, { "start": 2322, "end": 2336, "text": "(Fleiss, 1971)", "ref_id": "BIBREF13" } ], "ref_spans": [ { "start": 184, "end": 193, "text": "(Table 3)", "ref_id": "TABREF5" } ], "eq_spans": [], "section": "User Study on Plausibility", "sec_num": "4.6" }, { "text": "Faithfulness Many prior works have tried to improve the faithfulness of extractive rationales through the use of AAs (Bastings and Filippova, 2020) . Typically, this involves designing gradientbased (Sundararajan et al., 2017; Denil et al., 2014; Lundberg and Lee, 2017; Li et al., 2015) or perturbation-based (Li et al., 2016; Poerner et al., 2018; K\u00e1d\u00e1r et al., 2017) AAs. However, attribution algorithms cannot be optimized and tend to be compute-intensive (often requiring multiple LM forward/backward passes). Recently, Ismail et al. (2021) addressed the optimization issue by regularizing the task model to yield faithful rationales via the AA, while other works (Situ et al., 2021; Schwarzenberg et al., 2021) addressed the compute cost issue by training an LM (requiring only one forward pass) to mimic an AA's behavior. Another line of work aims to produce faithful rationales by construction, via SPPs (Jain et al., 2020; Yu et al., 2021; Bastings et al., 2019; Yu et al., 2019; Lei et al., 2016) . Still, SPPs' faithfulness can only guarantee sufficiency -not comprehensiveness (DeYoung et al., 2019) . Also, SPPs generally perform worse than vanilla LMs because they hide much of the original text input from the predictor and are hard to train end-to-end.", "cite_spans": [ { "start": 117, "end": 147, "text": "(Bastings and Filippova, 2020)", "ref_id": "BIBREF1" }, { "start": 199, "end": 226, "text": "(Sundararajan et al., 2017;", "ref_id": "BIBREF45" }, { "start": 227, "end": 246, "text": "Denil et al., 2014;", "ref_id": "BIBREF8" }, { "start": 247, "end": 270, "text": "Lundberg and Lee, 2017;", "ref_id": "BIBREF28" }, { "start": 271, "end": 287, "text": "Li et al., 2015)", "ref_id": "BIBREF24" }, { "start": 310, "end": 327, "text": "(Li et al., 2016;", "ref_id": "BIBREF25" }, { "start": 328, "end": 349, "text": "Poerner et al., 2018;", "ref_id": "BIBREF35" }, { "start": 350, "end": 369, "text": "K\u00e1d\u00e1r et al., 2017)", "ref_id": "BIBREF19" }, { "start": 525, "end": 545, "text": "Ismail et al. (2021)", "ref_id": "BIBREF16" }, { "start": 669, "end": 688, "text": "(Situ et al., 2021;", "ref_id": "BIBREF42" }, { "start": 689, "end": 716, "text": "Schwarzenberg et al., 2021)", "ref_id": "BIBREF39" }, { "start": 912, "end": 931, "text": "(Jain et al., 2020;", "ref_id": "BIBREF18" }, { "start": 932, "end": 948, "text": "Yu et al., 2021;", "ref_id": "BIBREF52" }, { "start": 949, "end": 971, "text": "Bastings et al., 2019;", "ref_id": "BIBREF0" }, { "start": 972, "end": 988, "text": "Yu et al., 2019;", "ref_id": "BIBREF51" }, { "start": 989, "end": 1006, "text": "Lei et al., 2016)", "ref_id": null }, { "start": 1089, "end": 1111, "text": "(DeYoung et al., 2019)", "ref_id": "BIBREF10" } ], "ref_spans": [], "eq_spans": [], "section": "Related Work", "sec_num": "5" }, { "text": "Plausibility Existing approaches for improving extractive rationale plausibility typically involve supervising LM-based extractors (Bhat et al., 2021) or SPPs (Jain et al., 2020; DeYoung et al., 2019) with gold rationales. However, existing LM-based extractors have not been trained for faithfulness, while SPPs' faithfulness by construction comes at the great cost of task performance. Meanwhile, more existing works focus on improving the plausibility of free-text rationales (Narang et al., 2020; Lakhotia et al., 2020; Camburu et al., 2018) , often with task-specific pipelines Kumar and Talukdar, 2020) .", "cite_spans": [ { "start": 131, "end": 150, "text": "(Bhat et al., 2021)", "ref_id": "BIBREF3" }, { "start": 159, "end": 178, "text": "(Jain et al., 2020;", "ref_id": "BIBREF18" }, { "start": 179, "end": 200, "text": "DeYoung et al., 2019)", "ref_id": "BIBREF10" }, { "start": 478, "end": 499, "text": "(Narang et al., 2020;", "ref_id": null }, { "start": 500, "end": 522, "text": "Lakhotia et al., 2020;", "ref_id": "BIBREF22" }, { "start": 523, "end": 544, "text": "Camburu et al., 2018)", "ref_id": "BIBREF4" }, { "start": 582, "end": 607, "text": "Kumar and Talukdar, 2020)", "ref_id": "BIBREF21" } ], "ref_spans": [], "eq_spans": [], "section": "Related Work", "sec_num": "5" }, { "text": "Connection to UNIREX Unlike prior works, UNIREX enables both the task model and rationale extractor to be jointly optimized for faithfulness, plausibility, and task performance. As a result, UNIREX-trained rationale extractors achieve a better balance of faithfulness and plausibility, without compromising the task model's performance. Also, by using a learned rationale extractor, which generally only requires one model forward pass, UNIREX does not have the computational expenses that limit many AAs.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Related Work", "sec_num": "5" }, { "text": "Here, we formalize the text classification problem in more detail.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "A.1 Text Classification", "sec_num": null }, { "text": "Let D = {X , Y} N i=1 be a dataset, where X = {x i } N i=1", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "A.1 Text Classification", "sec_num": null }, { "text": "are the text inputs, Y = {y * i } N i=1 are the labels, and N is the number of instances (x i , y * i ) in D. We also assume D can be partitioned into train set D train , dev set D dev , and test set D test . Let F task = f task (f enc (\u2022)) be a task LM, where f enc is the text encoder, and f task is the task output head. Typically, F task has a BERT-style architecture (Devlin et al., 2018) , in which f enc is a Transformer (Vaswani et al., 2017) while f task is a linear layer. Below, we define the sequence classification (SST, Movies, MultiRC, e-SNLI) and multi-choice QA (CoS-E) tasks, which are different types of text classification.", "cite_spans": [ { "start": 372, "end": 393, "text": "(Devlin et al., 2018)", "ref_id": "BIBREF9" }, { "start": 428, "end": 450, "text": "(Vaswani et al., 2017)", "ref_id": "BIBREF48" } ], "ref_spans": [], "eq_spans": [], "section": "A.1 Text Classification", "sec_num": null }, { "text": "Sequence Classification In sequence classification, x i is a token sequence (e.g., a single sentence, a pair of sentences), while y * i is the target class for x i . Here, we assume a fixed label space", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "A.1 Text Classification", "sec_num": null }, { "text": "Y = {1, ..., M } of size M , where y * i \u2208 Y for all i. Thus, f task outputs a vector of size M , such that F task (x i ) = f task (f enc (x i )) =\u0177 i \u2208 R M is the logit vector used to classify x i . Given\u0177 i = [\u0177 i,j ] M", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "A.1 Text Classification", "sec_num": null }, { "text": "j=1 , let y i = arg max j\u0177i,j be the class predicted by F task . The goal of sequence classification is to learn F task such that y * i = y i , for all (x i , y * i ) (Minaee et al., 2021) .", "cite_spans": [ { "start": 167, "end": 188, "text": "(Minaee et al., 2021)", "ref_id": "BIBREF31" } ], "ref_spans": [], "eq_spans": [], "section": "A.1 Text Classification", "sec_num": null }, { "text": "Multi-Choice QA Instead of a fixed label space, multi-choice QA has a different (but fixed-size) set of answer choices per instance. For instance i, let q i be the question (e.g., \"A friend is greeting me, what would they say?\") and", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "A.1 Text Classification", "sec_num": null }, { "text": "A i = {a i,j } M j=1", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "A.1 Text Classification", "sec_num": null }, { "text": "be the corresponding answer choices (e.g., {\"say hello\", \"greet\", \"associate\", \"socialize\", \"smile\"}), where M is now the number of answer choices. Define x i,j = q i \u2295 a i,j , where \u2295 denotes concatenation. In multi-choice QA, we have", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "A.1 Text Classification", "sec_num": null }, { "text": "x i = {x i,j } M j=1 , while y * i \u2208 A i", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "A.1 Text Classification", "sec_num": null }, { "text": "is the correct answer for x i . Thus, f task outputs a scalar, such that", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "A.1 Text Classification", "sec_num": null }, { "text": "F task (x i,j ) = f task (f enc (x i,j )) =\u0177 i,j \u2208 R is the logit for x i,j . Given\u0177 i = [\u0177 i,j ] M", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "A.1 Text Classification", "sec_num": null }, { "text": "j=1 , let j \u2032 = arg max j\u0177i,j , where y i = a i,j \u2032 is the answer predicted by F task . The goal of multi-choice QA is to learn F task such that y * i = y i , for all (x i , y * i ) (Talmor et al., 2018) .", "cite_spans": [ { "start": 182, "end": 203, "text": "(Talmor et al., 2018)", "ref_id": "BIBREF46" } ], "ref_spans": [], "eq_spans": [], "section": "A.1 Text Classification", "sec_num": null }, { "text": "A heuristic F task is an AA, which can be any handcrafted function that calculates an importance score s t i for each input token x t i (Bastings and Filippova, 2020) . AAs are typically gradient-based (Sundararajan et al., 2017; Denil et al., 2014; Lundberg and Lee, 2017; Li et al., 2015) or perturbationbased (Li et al., 2016; Poerner et al., 2018; K\u00e1d\u00e1r et al., 2017) methods. Gradient-based methods compute s t i via the gradient of F task 's output\u0177 i w.r.t. x t i , via one or more F task backward passes. Perturbation-based methods measure s t i as\u0177 i 's change when perturbing (e.g., removing) x t i , via multiple F task forward passes.", "cite_spans": [ { "start": 136, "end": 166, "text": "(Bastings and Filippova, 2020)", "ref_id": "BIBREF1" }, { "start": 202, "end": 229, "text": "(Sundararajan et al., 2017;", "ref_id": "BIBREF45" }, { "start": 230, "end": 249, "text": "Denil et al., 2014;", "ref_id": "BIBREF8" }, { "start": 250, "end": 273, "text": "Lundberg and Lee, 2017;", "ref_id": "BIBREF28" }, { "start": 274, "end": 290, "text": "Li et al., 2015)", "ref_id": "BIBREF24" }, { "start": 312, "end": 329, "text": "(Li et al., 2016;", "ref_id": "BIBREF25" }, { "start": 330, "end": 351, "text": "Poerner et al., 2018;", "ref_id": "BIBREF35" }, { "start": 352, "end": 371, "text": "K\u00e1d\u00e1r et al., 2017)", "ref_id": "BIBREF19" } ], "ref_spans": [], "eq_spans": [], "section": "A.2 Heuristic Rationale Extractors", "sec_num": null }, { "text": "AAs can be used out of the box without training and are designed to satisfy certain faithfulnessrelated axiomatic properties (Sundararajan et al., 2017; Lundberg and Lee, 2017) . However, AAs' lack of learnable parameters means they cannot be optimized for faithfulness/plausibility. Thus, if F task is trained for explainability using AA-based rationales, then only F task is optimized. Also, faithful AAs tend to be compute-intensive, requiring many F task backward/forward passes per instance (Sundararajan et al., 2017; Lundberg and Lee, 2017; Li et al., 2016) .", "cite_spans": [ { "start": 125, "end": 152, "text": "(Sundararajan et al., 2017;", "ref_id": "BIBREF45" }, { "start": 153, "end": 176, "text": "Lundberg and Lee, 2017)", "ref_id": "BIBREF28" }, { "start": 496, "end": 523, "text": "(Sundararajan et al., 2017;", "ref_id": "BIBREF45" }, { "start": 524, "end": 547, "text": "Lundberg and Lee, 2017;", "ref_id": "BIBREF28" }, { "start": 548, "end": 564, "text": "Li et al., 2016)", "ref_id": "BIBREF25" } ], "ref_spans": [], "eq_spans": [], "section": "A.2 Heuristic Rationale Extractors", "sec_num": null }, { "text": "If a learned rationale extractor is chosen, UNIREX enables users to specify how much gold rationale supervision to use. Ideally, each train instance would be annotated with a gold rationale. In this case, we could directly minimize the plausibility loss for each train instance. However, since gold rationales can be expensive to annotate, UNIREX provides a special batching procedure for training with limited gold rationale supervision.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "A.3 Gold Rationale Supervision", "sec_num": null }, { "text": "Given N train = |D train | train instances, let 0 < \u03b3 < 100 be the percentage of train instances with gold rationales, N gold = \u2308 \u03b3 100 N train \u2309 \u2265 1 be the number of train instances with gold rationales, b be the desired train batch size, and \u03b2 > 1 be a scaling factor. Define D gold \u2286 D train as the set of train instances with gold rationales, where |D gold | = N gold . Note that, if all train instances have gold rationales, then D gold = D train and \u03b3 = 100.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "A.3 Gold Rationale Supervision", "sec_num": null }, { "text": "Each batch is constructed as follows: (1) randomly sample b gold = max(1, b \u03b2 ) instances from D gold without replacement, then (2) randomly sample b \u2212 b gold instances from D train \\D gold without replacement. This results in a batch with b total train instances, b gold with gold rationales and the rest without. Since N gold is generally small, we only sample from D gold without replacement for a given batch, but not a given epoch. Thus, instances from D gold may appear more than once in the same epoch. However, we do sample from D train \\D gold without replacement for each batch and epoch, so every instance in D train \\D gold appears exactly once per epoch.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "A.3 Gold Rationale Supervision", "sec_num": null }, { "text": "After constructing the batch, we compute the plausibility loss for the batch as follows:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "A.3 Gold Rationale Supervision", "sec_num": null }, { "text": "b i=1 1 (x i ,y * i )\u2208D gold L plaus (F ext (x i ), r * i )", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "A.3 Gold Rationale Supervision", "sec_num": null }, { "text": ", where L plaus is the plausibility loss for train instance (x i , y * i ). This function zeroes out the plausibility loss for instances without gold rationales, so that plausibility is only being optimized with respect to instances with gold rationales. However, in Sec. ??, we show that it is possible to achieve high plausibility via rationale extractors trained on minimal gold rationale supervision.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "A.3 Gold Rationale Supervision", "sec_num": null }, { "text": "Sufficiency In addition, to the criteria presented in Sec. 3.2, we consider two other sufficiency loss functions. The first is the KL divergence criterion used in (Ismail et al., 2021) , which considers the entire label distribution and is defined as", "cite_spans": [ { "start": 163, "end": 184, "text": "(Ismail et al., 2021)", "ref_id": "BIBREF16" } ], "ref_spans": [], "eq_spans": [], "section": "A.4.1 Faithfulness", "sec_num": null }, { "text": "L suff-KL = KL(F task (r (k) i )) || F task (x i )).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "A.4.1 Faithfulness", "sec_num": null }, { "text": "The second is the mean absolute error (MAE) criterion, which is defined as", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "A.4.1 Faithfulness", "sec_num": null }, { "text": "L suff-MAE = |L CE (F task (r (k) i )), y * i ) \u2212 L CE (F task (x i ), y * i )|.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "A.4.1 Faithfulness", "sec_num": null }, { "text": "Unlike the difference criterion L suff-diff and margin criterion L suff-margin (Sec. 3.2), the MAE criterion assumes that using r (k) i as input should not yield better task performance than using x i as input. In our experiments, we find that L suff-margin is effective, though others (e.g., KL divergence, MAE) can be used too.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "A.4.1 Faithfulness", "sec_num": null }, { "text": "Similar to faithfulness, UNIREX places no restrictions on the choice of plausibility objective. As described in Sec. 3.2, given gold rationale r * i for input x i , plausibility optimization entails training F ext to predict binary importance label r * ,t i for each token x t i . This is essentially binary token classification, so one natural choice for L plaus is the token-level binary cross-entropy (BCE) criterion: L plaus-BCE = \u2212 t r * ,t i log(F ext (x t i )) (Sec. 3.2). Another option is the sequence-level KL divergence criterion, which is defined as:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "A.4.2 Plausibility", "sec_num": null }, { "text": "L plaus-KL = KL(F ext (x i ) || r * i )", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "A.4.2 Plausibility", "sec_num": null }, { "text": ". Additionally, we can directly penalize F ext (x i ) in the logit space via a linear loss, defined as:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "A.4.2 Plausibility", "sec_num": null }, { "text": "L plaus-linear = \u03a6(r * i ) F ext (x i ),", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "A.4.2 Plausibility", "sec_num": null }, { "text": "where \u03a6(u) = \u22122u + 1 maps positive and negative tokens to \u22121 and +1, respectively. The linear loss directly pushes the logits corresponding to positive/negative tokens to be higher/lower and increase the margin between them. To prevent linear loss values from becoming arbitrarily negative, we can also lower bound the loss with a margin m p , yielding:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "A.4.2 Plausibility", "sec_num": null }, { "text": "L plaus-linear-margin = max(\u2212m p , L plaus-linear ) + m p .", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "A.4.2 Plausibility", "sec_num": null }, { "text": "A.5 Implementation Details LM Architecture While many prior works use BERT (Devlin et al., 2018) Transformer LMs, BERT is limited to having sequences with up to 512 tokens, which is problematic since many datasets (e.g., Movies) contain much longer sequences. Meanwhile, BigBird (Zaheer et al., 2020) is a state-of-the-art Transformer LM designed to handle long input sequences with up to 4096 tokens. Thus, we use BigBird-Base, which is initialized with RoBERTa-Base (Liu et al., 2019) , in all of our experiments (i.e., both baselines and UNIREX). We obtain the pre-trained BigBird-Base model from the Hugging Face Transformers library (Wolf et al., 2019) . Note that UNIREX is agnostic to the choice of LM architecture, so RNNs, CNNs, and other Transformer LMs are also supported by UNIREX. However, we leave exploration of other LM architectures for future work.", "cite_spans": [ { "start": 75, "end": 96, "text": "(Devlin et al., 2018)", "ref_id": "BIBREF9" }, { "start": 279, "end": 300, "text": "(Zaheer et al., 2020)", "ref_id": "BIBREF53" }, { "start": 468, "end": 486, "text": "(Liu et al., 2019)", "ref_id": "BIBREF27" }, { "start": 638, "end": 657, "text": "(Wolf et al., 2019)", "ref_id": "BIBREF49" } ], "ref_spans": [], "eq_spans": [], "section": "A.4.2 Plausibility", "sec_num": null }, { "text": "Training Building upon Sec. ??, we discuss additional training details here. We find that \u03b1 c = 0.5 and \u03b1 s = 0.5 are usually best. For the batching factor \u03b2 (Sec. A.3), we use 2. For model selection, we choose the model with the best dev performance averaged over three seeds. We can also perform model selection based on dev explainability metrics, but we leave this extended tuning for future work. All experiments are implemented using PyTorch-Lightning (Paszke et al., 2019 ; Falcon and The PyTorch Lightning team, 2019). Fig. ?? shows the gold rationale data efficiency results for CoS-E, using the same setup as Sec. ??. Overall, we see that the CoS-E results are quite similar to the SST results. Again, UNIREX (DLM-FP) and UNIREX (SLM-FP) dominate across all \u03b3 values, with AUPRC slowly decreasing as \u03b3 de-creases. Interestingly, UNIREX (AA-FP) yields a noticeable dip in AUPRC for lower \u03b3 values. Since AA-FP has limited capacity (via the task model) for plausibility optimization, it is possible that this fluctuation is due to random noise. We leave further analysis of this for future work. ", "cite_spans": [ { "start": 458, "end": 478, "text": "(Paszke et al., 2019", "ref_id": "BIBREF34" } ], "ref_spans": [ { "start": 527, "end": 534, "text": "Fig. ??", "ref_id": null } ], "eq_spans": [], "section": "A.4.2 Plausibility", "sec_num": null }, { "text": "In this subsection, we present additional results from our experiments. Besides the aggregated results shown in Sec. 4 of the main text, Tables 4-10 contain more detailed results, using both raw and NRG metrics. Specifically , Tables 4-8 show all raw/NRG results for each dataset, Table 9 shows the ablation results for all raw metrics, and Table 10 includes the zero-shot explainability transfer results for UNIREX (SLM-FP). Generally, the computation of NRG should involve globally aggregating the raw metrics for all available methods, as done in the main results. However, for a number of more focused experiments (Tables 9-10), only a subset of the available methods are considered. Thus, to make the faithfulness results in Tables 9-10 easier to digest, we introduce a metric called Comp-Suff Difference (CSD), which locally aggregates comp and suff as: CSD = comp \u2212 suff. Therefore, since higher/lower comp/suff signal higher faithfulness, then higher CSD signals higher faithfulness. ", "cite_spans": [], "ref_spans": [ { "start": 225, "end": 237, "text": ", Tables 4-8", "ref_id": "TABREF7" }, { "start": 281, "end": 288, "text": "Table 9", "ref_id": null }, { "start": 341, "end": 349, "text": "Table 10", "ref_id": "TABREF1" } ], "eq_spans": [], "section": "A.7 Additional Empirical Results", "sec_num": null } ], "back_matter": [], "bib_entries": { "BIBREF0": { "ref_id": "b0", "title": "Interpretable neural predictions with differentiable binary variables", "authors": [ { "first": "Jasmijn", "middle": [], "last": "Bastings", "suffix": "" }, { "first": "Wilker", "middle": [], "last": "Aziz", "suffix": "" }, { "first": "Ivan", "middle": [], "last": "Titov", "suffix": "" } ], "year": 2019, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": { "arXiv": [ "arXiv:1905.08160" ] }, "num": null, "urls": [], "raw_text": "Jasmijn Bastings, Wilker Aziz, and Ivan Titov. 2019. Interpretable neural predictions with differentiable binary variables. arXiv preprint arXiv:1905.08160.", "links": null }, "BIBREF1": { "ref_id": "b1", "title": "The elephant in the interpretability room: Why use attention as explanation when we have saliency methods? arXiv preprint", "authors": [ { "first": "Jasmijn", "middle": [], "last": "Bastings", "suffix": "" }, { "first": "Katja", "middle": [], "last": "Filippova", "suffix": "" } ], "year": 2020, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": { "arXiv": [ "arXiv:2010.05607" ] }, "num": null, "urls": [], "raw_text": "Jasmijn Bastings and Katja Filippova. 2020. The ele- phant in the interpretability room: Why use atten- tion as explanation when we have saliency methods? arXiv preprint arXiv:2010.05607.", "links": null }, "BIBREF2": { "ref_id": "b2", "title": "On the dangers of stochastic parrots: Can language models be too big", "authors": [ { "first": "M", "middle": [], "last": "Emily", "suffix": "" }, { "first": "Timnit", "middle": [], "last": "Bender", "suffix": "" }, { "first": "Angelina", "middle": [], "last": "Gebru", "suffix": "" }, { "first": "Shmargaret", "middle": [], "last": "Mcmillan-Major", "suffix": "" }, { "first": "", "middle": [], "last": "Shmitchell", "suffix": "" } ], "year": 2021, "venue": "Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency", "volume": "", "issue": "", "pages": "610--623", "other_ids": {}, "num": null, "urls": [], "raw_text": "Emily M Bender, Timnit Gebru, Angelina McMillan- Major, and Shmargaret Shmitchell. 2021. On the dangers of stochastic parrots: Can language models be too big?. In Proceedings of the 2021 ACM Confer- ence on Fairness, Accountability, and Transparency, pages 610-623.", "links": null }, "BIBREF3": { "ref_id": "b3", "title": "Self-training with few-shot rationalization: Teacher explanations aid student in few-shot nlu", "authors": [ { "first": "Alessandro", "middle": [], "last": "Meghana Moorthy Bhat", "suffix": "" }, { "first": "Subhabrata", "middle": [], "last": "Sordoni", "suffix": "" }, { "first": "", "middle": [], "last": "Mukherjee", "suffix": "" } ], "year": 2021, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": { "arXiv": [ "arXiv:2109.08259" ] }, "num": null, "urls": [], "raw_text": "Meghana Moorthy Bhat, Alessandro Sordoni, and Sub- habrata Mukherjee. 2021. Self-training with few-shot rationalization: Teacher explanations aid student in few-shot nlu. arXiv preprint arXiv:2109.08259.", "links": null }, "BIBREF4": { "ref_id": "b4", "title": "Natural language inference with natural language explanations", "authors": [ { "first": "Oana-Maria", "middle": [], "last": "Camburu", "suffix": "" }, { "first": "Tim", "middle": [], "last": "Rockt\u00e4schel", "suffix": "" }, { "first": "Thomas", "middle": [], "last": "Lukasiewicz", "suffix": "" }, { "first": "Phil", "middle": [], "last": "Blunsom", "suffix": "" } ], "year": 2018, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": { "arXiv": [ "arXiv:1812.01193" ] }, "num": null, "urls": [], "raw_text": "Oana-Maria Camburu, Tim Rockt\u00e4schel, Thomas Lukasiewicz, and Phil Blunsom. 2018. e-snli: Natu- ral language inference with natural language expla- nations. arXiv preprint arXiv:1812.01193.", "links": null }, "BIBREF5": { "ref_id": "b5", "title": "Anirudh Rathore, and Chenhao Tan. 2020. Evaluating and characterizing human rationales", "authors": [ { "first": "Samuel", "middle": [], "last": "Carton", "suffix": "" } ], "year": null, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": { "arXiv": [ "arXiv:2010.04736" ] }, "num": null, "urls": [], "raw_text": "Samuel Carton, Anirudh Rathore, and Chenhao Tan. 2020. Evaluating and characterizing human ratio- nales. arXiv preprint arXiv:2010.04736.", "links": null }, "BIBREF6": { "ref_id": "b6", "title": "Salkg: Learning from knowledge graph explanations for commonsense reasoning", "authors": [ { "first": "Aaron", "middle": [], "last": "Chan", "suffix": "" }, { "first": "Jiashu", "middle": [], "last": "Xu", "suffix": "" }, { "first": "Boyuan", "middle": [], "last": "Long", "suffix": "" }, { "first": "Soumya", "middle": [], "last": "Sanyal", "suffix": "" }, { "first": "Tanishq", "middle": [], "last": "Gupta", "suffix": "" }, { "first": "Xiang", "middle": [], "last": "Ren", "suffix": "" } ], "year": 2021, "venue": "Advances in Neural Information Processing Systems", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Aaron Chan, Jiashu Xu, Boyuan Long, Soumya Sanyal, Tanishq Gupta, and Xiang Ren. 2021. Salkg: Learn- ing from knowledge graph explanations for common- sense reasoning. Advances in Neural Information Processing Systems, 34.", "links": null }, "BIBREF7": { "ref_id": "b7", "title": "Aitor Garc\u00eda-Pablos, and Montse Cuadros", "authors": [ { "first": "Ona", "middle": [], "last": "De Gibert", "suffix": "" }, { "first": "Naiara", "middle": [], "last": "Perez", "suffix": "" } ], "year": 2018, "venue": "Hate speech dataset from a white supremacy forum", "volume": "", "issue": "", "pages": "", "other_ids": { "arXiv": [ "arXiv:1809.04444" ] }, "num": null, "urls": [], "raw_text": "Ona de Gibert, Naiara Perez, Aitor Garc\u00eda-Pablos, and Montse Cuadros. 2018. Hate speech dataset from a white supremacy forum. arXiv preprint arXiv:1809.04444.", "links": null }, "BIBREF8": { "ref_id": "b8", "title": "Extraction of salient sentences from labelled documents", "authors": [ { "first": "Misha", "middle": [], "last": "Denil", "suffix": "" }, { "first": "Alban", "middle": [], "last": "Demiraj", "suffix": "" }, { "first": "Nando De", "middle": [], "last": "Freitas", "suffix": "" } ], "year": 2014, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": { "arXiv": [ "arXiv:1412.6815" ] }, "num": null, "urls": [], "raw_text": "Misha Denil, Alban Demiraj, and Nando De Freitas. 2014. Extraction of salient sentences from labelled documents. arXiv preprint arXiv:1412.6815.", "links": null }, "BIBREF9": { "ref_id": "b9", "title": "Bert: Pre-training of deep bidirectional transformers for language understanding", "authors": [ { "first": "Jacob", "middle": [], "last": "Devlin", "suffix": "" }, { "first": "Ming-Wei", "middle": [], "last": "Chang", "suffix": "" }, { "first": "Kenton", "middle": [], "last": "Lee", "suffix": "" }, { "first": "Kristina", "middle": [], "last": "Toutanova", "suffix": "" } ], "year": 2018, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": { "arXiv": [ "arXiv:1810.04805" ] }, "num": null, "urls": [], "raw_text": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2018. Bert: Pre-training of deep bidirectional transformers for language understand- ing. arXiv preprint arXiv:1810.04805.", "links": null }, "BIBREF10": { "ref_id": "b10", "title": "Eraser: A benchmark to evaluate rationalized nlp models", "authors": [ { "first": "Jay", "middle": [], "last": "Deyoung", "suffix": "" }, { "first": "Sarthak", "middle": [], "last": "Jain", "suffix": "" }, { "first": "Nazneen", "middle": [], "last": "Fatema Rajani", "suffix": "" }, { "first": "Eric", "middle": [], "last": "Lehman", "suffix": "" }, { "first": "Caiming", "middle": [], "last": "Xiong", "suffix": "" }, { "first": "Richard", "middle": [], "last": "Socher", "suffix": "" }, { "first": "Byron C", "middle": [], "last": "Wallace", "suffix": "" } ], "year": 2019, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": { "arXiv": [ "arXiv:1911.03429" ] }, "num": null, "urls": [], "raw_text": "Jay DeYoung, Sarthak Jain, Nazneen Fatema Rajani, Eric Lehman, Caiming Xiong, Richard Socher, and Byron C Wallace. 2019. Eraser: A benchmark to evaluate rationalized nlp models. arXiv preprint arXiv:1911.03429.", "links": null }, "BIBREF11": { "ref_id": "b11", "title": "Towards a rigorous science of interpretable machine learning", "authors": [ { "first": "Finale", "middle": [], "last": "Doshi", "suffix": "" }, { "first": "-", "middle": [], "last": "Velez", "suffix": "" }, { "first": "Been", "middle": [], "last": "Kim", "suffix": "" } ], "year": 2017, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": { "arXiv": [ "arXiv:1702.08608" ] }, "num": null, "urls": [], "raw_text": "Finale Doshi-Velez and Been Kim. 2017. Towards a rigorous science of interpretable machine learning. arXiv preprint arXiv:1702.08608.", "links": null }, "BIBREF12": { "ref_id": "b12", "title": "William Falcon and The PyTorch Lightning team", "authors": [], "year": 2019, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": { "DOI": [ "10.5281/zenodo.3828935" ] }, "num": null, "urls": [], "raw_text": "William Falcon and The PyTorch Lightning team. 2019. PyTorch Lightning.", "links": null }, "BIBREF13": { "ref_id": "b13", "title": "Measuring nominal scale agreement among many raters", "authors": [ { "first": "L", "middle": [], "last": "Joseph", "suffix": "" }, { "first": "", "middle": [], "last": "Fleiss", "suffix": "" } ], "year": 1971, "venue": "Psychological bulletin", "volume": "76", "issue": "5", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Joseph L Fleiss. 1971. Measuring nominal scale agree- ment among many raters. Psychological bulletin, 76(5):378.", "links": null }, "BIBREF14": { "ref_id": "b14", "title": "Evaluating explainable ai: Which algorithmic explanations help users predict model behavior?", "authors": [ { "first": "Peter", "middle": [], "last": "Hase", "suffix": "" }, { "first": "Mohit", "middle": [], "last": "Bansal", "suffix": "" } ], "year": 2020, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": { "arXiv": [ "arXiv:2005.01831" ] }, "num": null, "urls": [], "raw_text": "Peter Hase and Mohit Bansal. 2020. Evaluating ex- plainable ai: Which algorithmic explanations help users predict model behavior? arXiv preprint arXiv:2005.01831.", "links": null }, "BIBREF15": { "ref_id": "b15", "title": "A benchmark for interpretability methods in deep neural networks", "authors": [ { "first": "Sara", "middle": [], "last": "Hooker", "suffix": "" }, { "first": "Dumitru", "middle": [], "last": "Erhan", "suffix": "" }, { "first": "Pieter-Jan", "middle": [], "last": "Kindermans", "suffix": "" }, { "first": "Been", "middle": [], "last": "Kim", "suffix": "" } ], "year": 2018, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": { "arXiv": [ "arXiv:1806.10758" ] }, "num": null, "urls": [], "raw_text": "Sara Hooker, Dumitru Erhan, Pieter-Jan Kindermans, and Been Kim. 2018. A benchmark for interpretabil- ity methods in deep neural networks. arXiv preprint arXiv:1806.10758.", "links": null }, "BIBREF16": { "ref_id": "b16", "title": "Improving deep learning interpretability by saliency guided training", "authors": [ { "first": "Aya Abdelsalam", "middle": [], "last": "Ismail", "suffix": "" }, { "first": "Hector", "middle": [ "Corrada" ], "last": "Bravo", "suffix": "" }, { "first": "Soheil", "middle": [], "last": "Feizi", "suffix": "" } ], "year": 2021, "venue": "Advances in Neural Information Processing Systems", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Aya Abdelsalam Ismail, Hector Corrada Bravo, and Soheil Feizi. 2021. Improving deep learning inter- pretability by saliency guided training. Advances in Neural Information Processing Systems, 34.", "links": null }, "BIBREF17": { "ref_id": "b17", "title": "Towards faithfully interpretable nlp systems: How should we define and evaluate faithfulness?", "authors": [ { "first": "Alon", "middle": [], "last": "Jacovi", "suffix": "" }, { "first": "Yoav", "middle": [], "last": "Goldberg", "suffix": "" } ], "year": 2020, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": { "arXiv": [ "arXiv:2004.03685" ] }, "num": null, "urls": [], "raw_text": "Alon Jacovi and Yoav Goldberg. 2020. Towards faith- fully interpretable nlp systems: How should we define and evaluate faithfulness? arXiv preprint arXiv:2004.03685.", "links": null }, "BIBREF18": { "ref_id": "b18", "title": "Learning to faithfully rationalize by construction", "authors": [ { "first": "Sarthak", "middle": [], "last": "Jain", "suffix": "" }, { "first": "Sarah", "middle": [], "last": "Wiegreffe", "suffix": "" }, { "first": "Yuval", "middle": [], "last": "Pinter", "suffix": "" }, { "first": "Byron C", "middle": [], "last": "Wallace", "suffix": "" } ], "year": 2020, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": { "arXiv": [ "arXiv:2005.00115" ] }, "num": null, "urls": [], "raw_text": "Sarthak Jain, Sarah Wiegreffe, Yuval Pinter, and Byron C Wallace. 2020. Learning to faith- fully rationalize by construction. arXiv preprint arXiv:2005.00115.", "links": null }, "BIBREF19": { "ref_id": "b19", "title": "Representation of linguistic form and function in recurrent neural networks", "authors": [ { "first": "Akos", "middle": [], "last": "K\u00e1d\u00e1r", "suffix": "" }, { "first": "Grzegorz", "middle": [], "last": "Chrupa\u0142a", "suffix": "" }, { "first": "Afra", "middle": [], "last": "Alishahi", "suffix": "" } ], "year": 2017, "venue": "Computational Linguistics", "volume": "43", "issue": "4", "pages": "761--780", "other_ids": {}, "num": null, "urls": [], "raw_text": "Akos K\u00e1d\u00e1r, Grzegorz Chrupa\u0142a, and Afra Alishahi. 2017. Representation of linguistic form and func- tion in recurrent neural networks. Computational Linguistics, 43(4):761-780.", "links": null }, "BIBREF20": { "ref_id": "b20", "title": "Looking beyond the surface: A challenge set for reading comprehension over multiple sentences", "authors": [ { "first": "Daniel", "middle": [], "last": "Khashabi", "suffix": "" }, { "first": "Snigdha", "middle": [], "last": "Chaturvedi", "suffix": "" }, { "first": "Michael", "middle": [], "last": "Roth", "suffix": "" }, { "first": "Shyam", "middle": [], "last": "Upadhyay", "suffix": "" }, { "first": "Dan", "middle": [], "last": "Roth", "suffix": "" } ], "year": 2018, "venue": "Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies", "volume": "1", "issue": "", "pages": "252--262", "other_ids": {}, "num": null, "urls": [], "raw_text": "Daniel Khashabi, Snigdha Chaturvedi, Michael Roth, Shyam Upadhyay, and Dan Roth. 2018. Looking beyond the surface: A challenge set for reading com- prehension over multiple sentences. In Proceedings of the 2018 Conference of the North American Chap- ter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Pa- pers), pages 252-262.", "links": null }, "BIBREF21": { "ref_id": "b21", "title": "Nile: Natural language inference with faithful natural language explanations", "authors": [ { "first": "Sawan", "middle": [], "last": "Kumar", "suffix": "" }, { "first": "Partha", "middle": [], "last": "Talukdar", "suffix": "" } ], "year": 2020, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": { "arXiv": [ "arXiv:2005.12116" ] }, "num": null, "urls": [], "raw_text": "Sawan Kumar and Partha Talukdar. 2020. Nile: Natu- ral language inference with faithful natural language explanations. arXiv preprint arXiv:2005.12116.", "links": null }, "BIBREF22": { "ref_id": "b22", "title": "Fid-ex: Improving sequence-to-sequence models for extractive rationale generation", "authors": [ { "first": "Kushal", "middle": [], "last": "Lakhotia", "suffix": "" }, { "first": "Bhargavi", "middle": [], "last": "Paranjape", "suffix": "" }, { "first": "Asish", "middle": [], "last": "Ghoshal", "suffix": "" }, { "first": "Wen-Tau", "middle": [], "last": "Yih", "suffix": "" }, { "first": "Yashar", "middle": [], "last": "Mehdad", "suffix": "" }, { "first": "Srinivasan", "middle": [], "last": "Iyer", "suffix": "" } ], "year": 2020, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": { "arXiv": [ "arXiv:2012.15482" ] }, "num": null, "urls": [], "raw_text": "Kushal Lakhotia, Bhargavi Paranjape, Asish Ghoshal, Wen-tau Yih, Yashar Mehdad, and Srinivasan Iyer. 2020. Fid-ex: Improving sequence-to-sequence mod- els for extractive rationale generation. arXiv preprint arXiv:2012.15482.", "links": null }, "BIBREF24": { "ref_id": "b24", "title": "Visualizing and understanding neural models in nlp", "authors": [ { "first": "Jiwei", "middle": [], "last": "Li", "suffix": "" }, { "first": "Xinlei", "middle": [], "last": "Chen", "suffix": "" }, { "first": "Eduard", "middle": [], "last": "Hovy", "suffix": "" }, { "first": "Dan", "middle": [], "last": "Jurafsky", "suffix": "" } ], "year": 2015, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": { "arXiv": [ "arXiv:1506.01066" ] }, "num": null, "urls": [], "raw_text": "Jiwei Li, Xinlei Chen, Eduard Hovy, and Dan Jurafsky. 2015. Visualizing and understanding neural models in nlp. arXiv preprint arXiv:1506.01066.", "links": null }, "BIBREF25": { "ref_id": "b25", "title": "Understanding neural networks through representation erasure", "authors": [ { "first": "Jiwei", "middle": [], "last": "Li", "suffix": "" }, { "first": "Will", "middle": [], "last": "Monroe", "suffix": "" }, { "first": "Dan", "middle": [], "last": "Jurafsky", "suffix": "" } ], "year": 2016, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": { "arXiv": [ "arXiv:1612.08220" ] }, "num": null, "urls": [], "raw_text": "Jiwei Li, Will Monroe, and Dan Jurafsky. 2016. Un- derstanding neural networks through representation erasure. arXiv preprint arXiv:1612.08220.", "links": null }, "BIBREF26": { "ref_id": "b26", "title": "The mythos of model interpretability: In machine learning, the concept of interpretability is both important and slippery", "authors": [ { "first": "", "middle": [], "last": "Zachary C Lipton", "suffix": "" } ], "year": 2018, "venue": "Queue", "volume": "16", "issue": "3", "pages": "31--57", "other_ids": {}, "num": null, "urls": [], "raw_text": "Zachary C Lipton. 2018. The mythos of model inter- pretability: In machine learning, the concept of in- terpretability is both important and slippery. Queue, 16(3):31-57.", "links": null }, "BIBREF27": { "ref_id": "b27", "title": "Roberta: A robustly optimized bert pretraining approach", "authors": [ { "first": "Yinhan", "middle": [], "last": "Liu", "suffix": "" }, { "first": "Myle", "middle": [], "last": "Ott", "suffix": "" }, { "first": "Naman", "middle": [], "last": "Goyal", "suffix": "" }, { "first": "Jingfei", "middle": [], "last": "Du", "suffix": "" }, { "first": "Mandar", "middle": [], "last": "Joshi", "suffix": "" }, { "first": "Danqi", "middle": [], "last": "Chen", "suffix": "" }, { "first": "Omer", "middle": [], "last": "Levy", "suffix": "" }, { "first": "Mike", "middle": [], "last": "Lewis", "suffix": "" }, { "first": "Luke", "middle": [], "last": "Zettlemoyer", "suffix": "" }, { "first": "Veselin", "middle": [], "last": "Stoyanov", "suffix": "" } ], "year": 2019, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": { "arXiv": [ "arXiv:1907.11692" ] }, "num": null, "urls": [], "raw_text": "Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Man- dar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A robustly optimized bert pretraining ap- proach. arXiv preprint arXiv:1907.11692.", "links": null }, "BIBREF28": { "ref_id": "b28", "title": "A unified approach to interpreting model predictions", "authors": [ { "first": "M", "middle": [], "last": "Scott", "suffix": "" }, { "first": "Su-In", "middle": [], "last": "Lundberg", "suffix": "" }, { "first": "", "middle": [], "last": "Lee", "suffix": "" } ], "year": 2017, "venue": "Proceedings of the 31st international conference on neural information processing systems", "volume": "", "issue": "", "pages": "4768--4777", "other_ids": {}, "num": null, "urls": [], "raw_text": "Scott M Lundberg and Su-In Lee. 2017. A unified ap- proach to interpreting model predictions. In Proceed- ings of the 31st international conference on neural information processing systems, pages 4768-4777.", "links": null }, "BIBREF29": { "ref_id": "b29", "title": "Local interpretations for explainable natural language processing: A survey", "authors": [ { "first": "Siwen", "middle": [], "last": "Luo", "suffix": "" }, { "first": "Hamish", "middle": [], "last": "Ivison", "suffix": "" }, { "first": "Caren", "middle": [], "last": "Han", "suffix": "" }, { "first": "Josiah", "middle": [], "last": "Poon", "suffix": "" } ], "year": 2021, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": { "arXiv": [ "arXiv:2103.11072" ] }, "num": null, "urls": [], "raw_text": "Siwen Luo, Hamish Ivison, Caren Han, and Josiah Poon. 2021. Local interpretations for explainable natu- ral language processing: A survey. arXiv preprint arXiv:2103.11072.", "links": null }, "BIBREF30": { "ref_id": "b30", "title": "Hidden factors and hidden topics: understanding rating dimensions with review text", "authors": [ { "first": "Julian", "middle": [], "last": "Mcauley", "suffix": "" }, { "first": "Jure", "middle": [], "last": "Leskovec", "suffix": "" } ], "year": 2013, "venue": "Proceedings of the 7th ACM conference on Recommender systems", "volume": "", "issue": "", "pages": "165--172", "other_ids": {}, "num": null, "urls": [], "raw_text": "Julian McAuley and Jure Leskovec. 2013. Hidden fac- tors and hidden topics: understanding rating dimen- sions with review text. In Proceedings of the 7th ACM conference on Recommender systems, pages 165-172.", "links": null }, "BIBREF31": { "ref_id": "b31", "title": "Deep learning-based text classification: A comprehensive review", "authors": [ { "first": "Shervin", "middle": [], "last": "Minaee", "suffix": "" }, { "first": "Nal", "middle": [], "last": "Kalchbrenner", "suffix": "" }, { "first": "Erik", "middle": [], "last": "Cambria", "suffix": "" }, { "first": "Narjes", "middle": [], "last": "Nikzad", "suffix": "" }, { "first": "Meysam", "middle": [], "last": "Chenaghlu", "suffix": "" }, { "first": "Jianfeng", "middle": [], "last": "Gao", "suffix": "" } ], "year": 2021, "venue": "ACM Computing Surveys (CSUR)", "volume": "54", "issue": "3", "pages": "1--40", "other_ids": {}, "num": null, "urls": [], "raw_text": "Shervin Minaee, Nal Kalchbrenner, Erik Cambria, Nar- jes Nikzad, Meysam Chenaghlu, and Jianfeng Gao. 2021. Deep learning-based text classification: A comprehensive review. ACM Computing Surveys (CSUR), 54(3):1-40.", "links": null }, "BIBREF32": { "ref_id": "b32", "title": "Noah Fiedel, and Karishma Malkan. 2020. Wt5?! training text-to-text models to explain their predictions", "authors": [ { "first": "Sharan", "middle": [], "last": "Narang", "suffix": "" }, { "first": "Colin", "middle": [], "last": "Raffel", "suffix": "" }, { "first": "Katherine", "middle": [], "last": "Lee", "suffix": "" }, { "first": "Adam", "middle": [], "last": "Roberts", "suffix": "" } ], "year": null, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": { "arXiv": [ "arXiv:2004.14546" ] }, "num": null, "urls": [], "raw_text": "Sharan Narang, Colin Raffel, Katherine Lee, Adam Roberts, Noah Fiedel, and Karishma Malkan. 2020. Wt5?! training text-to-text models to explain their predictions. arXiv preprint arXiv:2004.14546.", "links": null }, "BIBREF33": { "ref_id": "b33", "title": "An information bottleneck approach for controlling conciseness in rationale extraction", "authors": [ { "first": "Mandar", "middle": [], "last": "Bhargavi Paranjape", "suffix": "" }, { "first": "John", "middle": [], "last": "Joshi", "suffix": "" }, { "first": "Hannaneh", "middle": [], "last": "Thickstun", "suffix": "" }, { "first": "Luke", "middle": [], "last": "Hajishirzi", "suffix": "" }, { "first": "", "middle": [], "last": "Zettlemoyer", "suffix": "" } ], "year": 2020, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": { "arXiv": [ "arXiv:2005.00652" ] }, "num": null, "urls": [], "raw_text": "Bhargavi Paranjape, Mandar Joshi, John Thickstun, Hannaneh Hajishirzi, and Luke Zettlemoyer. 2020. An information bottleneck approach for controlling conciseness in rationale extraction. arXiv preprint arXiv:2005.00652.", "links": null }, "BIBREF34": { "ref_id": "b34", "title": "Pytorch: An imperative style, high-performance deep learning library", "authors": [ { "first": "Adam", "middle": [], "last": "Paszke", "suffix": "" }, { "first": "Sam", "middle": [], "last": "Gross", "suffix": "" }, { "first": "Francisco", "middle": [], "last": "Massa", "suffix": "" }, { "first": "Adam", "middle": [], "last": "Lerer", "suffix": "" }, { "first": "James", "middle": [], "last": "Bradbury", "suffix": "" }, { "first": "Gregory", "middle": [], "last": "Chanan", "suffix": "" }, { "first": "Trevor", "middle": [], "last": "Killeen", "suffix": "" }, { "first": "Zeming", "middle": [], "last": "Lin", "suffix": "" }, { "first": "Natalia", "middle": [], "last": "Gimelshein", "suffix": "" }, { "first": "Luca", "middle": [], "last": "Antiga", "suffix": "" } ], "year": 2019, "venue": "Advances in neural information processing systems", "volume": "32", "issue": "", "pages": "8026--8037", "other_ids": {}, "num": null, "urls": [], "raw_text": "Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. 2019. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32:8026- 8037.", "links": null }, "BIBREF35": { "ref_id": "b35", "title": "Evaluating neural network explanation methods using hybrid documents and morphological agreement", "authors": [ { "first": "Nina", "middle": [], "last": "Poerner", "suffix": "" }, { "first": "Benjamin", "middle": [], "last": "Roth", "suffix": "" }, { "first": "Hinrich", "middle": [], "last": "Sch\u00fctze", "suffix": "" } ], "year": 2018, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": { "arXiv": [ "arXiv:1801.06422" ] }, "num": null, "urls": [], "raw_text": "Nina Poerner, Benjamin Roth, and Hinrich Sch\u00fctze. 2018. Evaluating neural network explanation meth- ods using hybrid documents and morphological agreement. arXiv preprint arXiv:1801.06422.", "links": null }, "BIBREF36": { "ref_id": "b36", "title": "Evaluating explanations: How much do explanations from the teacher aid students?", "authors": [ { "first": "Danish", "middle": [], "last": "Pruthi", "suffix": "" }, { "first": "Bhuwan", "middle": [], "last": "Dhingra", "suffix": "" }, { "first": "Livio Baldini", "middle": [], "last": "Soares", "suffix": "" }, { "first": "Michael", "middle": [], "last": "Collins", "suffix": "" }, { "first": "C", "middle": [], "last": "Zachary", "suffix": "" }, { "first": "Graham", "middle": [], "last": "Lipton", "suffix": "" }, { "first": "William", "middle": [ "W" ], "last": "Neubig", "suffix": "" }, { "first": "", "middle": [], "last": "Cohen", "suffix": "" } ], "year": 2020, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": { "arXiv": [ "arXiv:2012.00893" ] }, "num": null, "urls": [], "raw_text": "Danish Pruthi, Bhuwan Dhingra, Livio Baldini Soares, Michael Collins, Zachary C Lipton, Graham Neubig, and William W Cohen. 2020. Evaluating explana- tions: How much do explanations from the teacher aid students? arXiv preprint arXiv:2012.00893.", "links": null }, "BIBREF37": { "ref_id": "b37", "title": "Explain yourself! leveraging language models for commonsense reasoning", "authors": [ { "first": "Bryan", "middle": [], "last": "Nazneen Fatema Rajani", "suffix": "" }, { "first": "Caiming", "middle": [], "last": "Mccann", "suffix": "" }, { "first": "Richard", "middle": [], "last": "Xiong", "suffix": "" }, { "first": "", "middle": [], "last": "Socher", "suffix": "" } ], "year": 2019, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": { "arXiv": [ "arXiv:1906.02361" ] }, "num": null, "urls": [], "raw_text": "Nazneen Fatema Rajani, Bryan McCann, Caiming Xiong, and Richard Socher. 2019. Explain your- self! leveraging language models for commonsense reasoning. arXiv preprint arXiv:1906.02361.", "links": null }, "BIBREF38": { "ref_id": "b38", "title": "Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead", "authors": [ { "first": "Cynthia", "middle": [], "last": "Rudin", "suffix": "" } ], "year": 2019, "venue": "Nature Machine Intelligence", "volume": "1", "issue": "5", "pages": "206--215", "other_ids": {}, "num": null, "urls": [], "raw_text": "Cynthia Rudin. 2019. Stop explaining black box ma- chine learning models for high stakes decisions and use interpretable models instead. Nature Machine Intelligence, 1(5):206-215.", "links": null }, "BIBREF39": { "ref_id": "b39", "title": "Efficient explanations from empirical explainers", "authors": [ { "first": "Robert", "middle": [], "last": "Schwarzenberg", "suffix": "" }, { "first": "Nils", "middle": [], "last": "Feldhus", "suffix": "" }, { "first": "Sebastian", "middle": [], "last": "M\u00f6ller", "suffix": "" } ], "year": 2021, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": { "arXiv": [ "arXiv:2103.15429" ] }, "num": null, "urls": [], "raw_text": "Robert Schwarzenberg, Nils Feldhus, and Sebastian M\u00f6ller. 2021. Efficient explanations from empirical explainers. arXiv preprint arXiv:2103.15429.", "links": null }, "BIBREF40": { "ref_id": "b40", "title": "Learning important features through propagating activation differences", "authors": [ { "first": "Avanti", "middle": [], "last": "Shrikumar", "suffix": "" }, { "first": "Peyton", "middle": [], "last": "Greenside", "suffix": "" }, { "first": "Anshul", "middle": [], "last": "Kundaje", "suffix": "" } ], "year": 2017, "venue": "International Conference on Machine Learning", "volume": "", "issue": "", "pages": "3145--3153", "other_ids": {}, "num": null, "urls": [], "raw_text": "Avanti Shrikumar, Peyton Greenside, and Anshul Kun- daje. 2017. Learning important features through propagating activation differences. In International Conference on Machine Learning, pages 3145-3153. PMLR.", "links": null }, "BIBREF41": { "ref_id": "b41", "title": "Deep inside convolutional networks: Visualising image classification models and saliency maps", "authors": [ { "first": "Karen", "middle": [], "last": "Simonyan", "suffix": "" }, { "first": "Andrea", "middle": [], "last": "Vedaldi", "suffix": "" }, { "first": "Andrew", "middle": [], "last": "Zisserman", "suffix": "" } ], "year": 2013, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": { "arXiv": [ "arXiv:1312.6034" ] }, "num": null, "urls": [], "raw_text": "Karen Simonyan, Andrea Vedaldi, and Andrew Zis- serman. 2013. Deep inside convolutional networks: Visualising image classification models and saliency maps. arXiv preprint arXiv:1312.6034.", "links": null }, "BIBREF42": { "ref_id": "b42", "title": "Learning to explain: Generating stable explanations fast", "authors": [ { "first": "Xuelin", "middle": [], "last": "Situ", "suffix": "" }, { "first": "Ingrid", "middle": [], "last": "Zukerman", "suffix": "" }, { "first": "Cecile", "middle": [], "last": "Paris", "suffix": "" }, { "first": "Sameen", "middle": [], "last": "Maruf", "suffix": "" }, { "first": "Gholamreza", "middle": [], "last": "Haffari", "suffix": "" } ], "year": 2021, "venue": "Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing", "volume": "1", "issue": "", "pages": "5340--5355", "other_ids": {}, "num": null, "urls": [], "raw_text": "Xuelin Situ, Ingrid Zukerman, Cecile Paris, Sameen Maruf, and Gholamreza Haffari. 2021. Learning to explain: Generating stable explanations fast. In Proceedings of the 59th Annual Meeting of the Asso- ciation for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 5340- 5355.", "links": null }, "BIBREF43": { "ref_id": "b43", "title": "Recursive deep models for semantic compositionality over a sentiment treebank", "authors": [ { "first": "Richard", "middle": [], "last": "Socher", "suffix": "" }, { "first": "Alex", "middle": [], "last": "Perelygin", "suffix": "" }, { "first": "Jean", "middle": [], "last": "Wu", "suffix": "" }, { "first": "Jason", "middle": [], "last": "Chuang", "suffix": "" }, { "first": "D", "middle": [], "last": "Christopher", "suffix": "" }, { "first": "", "middle": [], "last": "Manning", "suffix": "" }, { "first": "Y", "middle": [], "last": "Andrew", "suffix": "" }, { "first": "Christopher", "middle": [], "last": "Ng", "suffix": "" }, { "first": "", "middle": [], "last": "Potts", "suffix": "" } ], "year": 2013, "venue": "Proceedings of the 2013 conference on empirical methods in natural language processing", "volume": "", "issue": "", "pages": "1631--1642", "other_ids": {}, "num": null, "urls": [], "raw_text": "Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D Manning, Andrew Y Ng, and Christopher Potts. 2013. Recursive deep models for semantic compositionality over a sentiment treebank. In Proceedings of the 2013 conference on empiri- cal methods in natural language processing, pages 1631-1642.", "links": null }, "BIBREF44": { "ref_id": "b44", "title": "Do human rationales improve machine explanations? arXiv preprint", "authors": [ { "first": "Julia", "middle": [], "last": "Strout", "suffix": "" }, { "first": "Ye", "middle": [], "last": "Zhang", "suffix": "" }, { "first": "Raymond J", "middle": [], "last": "Mooney", "suffix": "" } ], "year": 2019, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": { "arXiv": [ "arXiv:1905.13714" ] }, "num": null, "urls": [], "raw_text": "Julia Strout, Ye Zhang, and Raymond J Mooney. 2019. Do human rationales improve machine explanations? arXiv preprint arXiv:1905.13714.", "links": null }, "BIBREF45": { "ref_id": "b45", "title": "Axiomatic attribution for deep networks", "authors": [ { "first": "Mukund", "middle": [], "last": "Sundararajan", "suffix": "" }, { "first": "Ankur", "middle": [], "last": "Taly", "suffix": "" }, { "first": "Qiqi", "middle": [], "last": "Yan", "suffix": "" } ], "year": 2017, "venue": "International Conference on Machine Learning", "volume": "", "issue": "", "pages": "3319--3328", "other_ids": {}, "num": null, "urls": [], "raw_text": "Mukund Sundararajan, Ankur Taly, and Qiqi Yan. 2017. Axiomatic attribution for deep networks. In Inter- national Conference on Machine Learning, pages 3319-3328. PMLR.", "links": null }, "BIBREF46": { "ref_id": "b46", "title": "Commonsenseqa: A question answering challenge targeting commonsense knowledge", "authors": [ { "first": "Alon", "middle": [], "last": "Talmor", "suffix": "" }, { "first": "Jonathan", "middle": [], "last": "Herzig", "suffix": "" }, { "first": "Nicholas", "middle": [], "last": "Lourie", "suffix": "" }, { "first": "Jonathan", "middle": [], "last": "Berant", "suffix": "" } ], "year": 2018, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": { "arXiv": [ "arXiv:1811.00937" ] }, "num": null, "urls": [], "raw_text": "Alon Talmor, Jonathan Herzig, Nicholas Lourie, and Jonathan Berant. 2018. Commonsenseqa: A question answering challenge targeting commonsense knowl- edge. arXiv preprint arXiv:1811.00937.", "links": null }, "BIBREF47": { "ref_id": "b47", "title": "Semeval-2018 task 3: Irony detection in english tweets", "authors": [ { "first": "Cynthia", "middle": [], "last": "Van Hee", "suffix": "" }, { "first": "Els", "middle": [], "last": "Lefever", "suffix": "" }, { "first": "V\u00e9ronique", "middle": [], "last": "Hoste", "suffix": "" } ], "year": 2018, "venue": "Proceedings of The 12th International Workshop on Semantic Evaluation", "volume": "", "issue": "", "pages": "39--50", "other_ids": {}, "num": null, "urls": [], "raw_text": "Cynthia Van Hee, Els Lefever, and V\u00e9ronique Hoste. 2018. Semeval-2018 task 3: Irony detection in en- glish tweets. In Proceedings of The 12th Interna- tional Workshop on Semantic Evaluation, pages 39- 50.", "links": null }, "BIBREF48": { "ref_id": "b48", "title": "Attention is all you need", "authors": [ { "first": "Ashish", "middle": [], "last": "Vaswani", "suffix": "" }, { "first": "Noam", "middle": [], "last": "Shazeer", "suffix": "" }, { "first": "Niki", "middle": [], "last": "Parmar", "suffix": "" }, { "first": "Jakob", "middle": [], "last": "Uszkoreit", "suffix": "" }, { "first": "Llion", "middle": [], "last": "Jones", "suffix": "" }, { "first": "Aidan", "middle": [ "N" ], "last": "Gomez", "suffix": "" }, { "first": "\u0141ukasz", "middle": [], "last": "Kaiser", "suffix": "" }, { "first": "Illia", "middle": [], "last": "Polosukhin", "suffix": "" } ], "year": 2017, "venue": "Advances in neural information processing systems", "volume": "", "issue": "", "pages": "5998--6008", "other_ids": {}, "num": null, "urls": [], "raw_text": "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, \u0141ukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Advances in neural information pro- cessing systems, pages 5998-6008.", "links": null }, "BIBREF49": { "ref_id": "b49", "title": "Huggingface's transformers: State-ofthe-art natural language processing", "authors": [ { "first": "Thomas", "middle": [], "last": "Wolf", "suffix": "" }, { "first": "Lysandre", "middle": [], "last": "Debut", "suffix": "" }, { "first": "Victor", "middle": [], "last": "Sanh", "suffix": "" }, { "first": "Julien", "middle": [], "last": "Chaumond", "suffix": "" }, { "first": "Clement", "middle": [], "last": "Delangue", "suffix": "" }, { "first": "Anthony", "middle": [], "last": "Moi", "suffix": "" }, { "first": "Pierric", "middle": [], "last": "Cistac", "suffix": "" }, { "first": "Tim", "middle": [], "last": "Rault", "suffix": "" }, { "first": "R\u00e9mi", "middle": [], "last": "Louf", "suffix": "" }, { "first": "Morgan", "middle": [], "last": "Funtowicz", "suffix": "" } ], "year": 2019, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": { "arXiv": [ "arXiv:1910.03771" ] }, "num": null, "urls": [], "raw_text": "Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pier- ric Cistac, Tim Rault, R\u00e9mi Louf, Morgan Funtowicz, et al. 2019. Huggingface's transformers: State-of- the-art natural language processing. arXiv preprint arXiv:1910.03771.", "links": null }, "BIBREF50": { "ref_id": "b50", "title": "Crossfit: A few-shot learning challenge for cross-task generalization in nlp", "authors": [ { "first": "Qinyuan", "middle": [], "last": "Ye", "suffix": "" }, { "first": "Xiang", "middle": [], "last": "Bill Yuchen Lin", "suffix": "" }, { "first": "", "middle": [], "last": "Ren", "suffix": "" } ], "year": 2021, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": { "arXiv": [ "arXiv:2104.08835" ] }, "num": null, "urls": [], "raw_text": "Qinyuan Ye, Bill Yuchen Lin, and Xiang Ren. 2021. Crossfit: A few-shot learning challenge for cross-task generalization in nlp. arXiv preprint arXiv:2104.08835.", "links": null }, "BIBREF51": { "ref_id": "b51", "title": "Rethinking cooperative rationalization: Introspective extraction and complement control", "authors": [ { "first": "Mo", "middle": [], "last": "Yu", "suffix": "" }, { "first": "Shiyu", "middle": [], "last": "Chang", "suffix": "" }, { "first": "Yang", "middle": [], "last": "Zhang", "suffix": "" }, { "first": "Tommi", "middle": [ "S" ], "last": "Jaakkola", "suffix": "" } ], "year": 2019, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": { "arXiv": [ "arXiv:1910.13294" ] }, "num": null, "urls": [], "raw_text": "Mo Yu, Shiyu Chang, Yang Zhang, and Tommi S Jaakkola. 2019. Rethinking cooperative rationaliza- tion: Introspective extraction and complement con- trol. arXiv preprint arXiv:1910.13294.", "links": null }, "BIBREF52": { "ref_id": "b52", "title": "Understanding interlocking dynamics of cooperative rationalization", "authors": [ { "first": "Mo", "middle": [], "last": "Yu", "suffix": "" }, { "first": "Yang", "middle": [], "last": "Zhang", "suffix": "" }, { "first": "Shiyu", "middle": [], "last": "Chang", "suffix": "" }, { "first": "Tommi", "middle": [], "last": "Jaakkola", "suffix": "" } ], "year": 2021, "venue": "Advances in Neural Information Processing Systems", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Mo Yu, Yang Zhang, Shiyu Chang, and Tommi Jaakkola. 2021. Understanding interlocking dynamics of coop- erative rationalization. Advances in Neural Informa- tion Processing Systems, 34.", "links": null }, "BIBREF53": { "ref_id": "b53", "title": "Big bird: Transformers for longer sequences", "authors": [ { "first": "Manzil", "middle": [], "last": "Zaheer", "suffix": "" }, { "first": "Guru", "middle": [], "last": "Guruganesh", "suffix": "" }, { "first": "Joshua", "middle": [], "last": "Kumar Avinava Dubey", "suffix": "" }, { "first": "Chris", "middle": [], "last": "Ainslie", "suffix": "" }, { "first": "Santiago", "middle": [], "last": "Alberti", "suffix": "" }, { "first": "Philip", "middle": [], "last": "Ontanon", "suffix": "" }, { "first": "Anirudh", "middle": [], "last": "Pham", "suffix": "" }, { "first": "Qifan", "middle": [], "last": "Ravula", "suffix": "" }, { "first": "Li", "middle": [], "last": "Wang", "suffix": "" }, { "first": "", "middle": [], "last": "Yang", "suffix": "" } ], "year": 2020, "venue": "NeurIPS", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Manzil Zaheer, Guru Guruganesh, Kumar Avinava Dubey, Joshua Ainslie, Chris Alberti, Santiago On- tanon, Philip Pham, Anirudh Ravula, Qifan Wang, Li Yang, et al. 2020. Big bird: Transformers for longer sequences. In NeurIPS.", "links": null }, "BIBREF54": { "ref_id": "b54", "title": "Modeling annotators: A generative approach to learning from annotator rationales", "authors": [ { "first": "Omar", "middle": [], "last": "Zaidan", "suffix": "" }, { "first": "Jason", "middle": [], "last": "Eisner", "suffix": "" } ], "year": 2008, "venue": "Proceedings of the 2008 conference on Empirical methods in natural language processing", "volume": "", "issue": "", "pages": "31--40", "other_ids": {}, "num": null, "urls": [], "raw_text": "Omar Zaidan and Jason Eisner. 2008. Modeling an- notators: A generative approach to learning from annotator rationales. In Proceedings of the 2008 con- ference on Empirical methods in natural language processing, pages 31-40.", "links": null }, "BIBREF55": { "ref_id": "b55", "title": "Semeval-2019 task 6: Identifying and categorizing offensive language in social media (offenseval)", "authors": [ { "first": "Marcos", "middle": [], "last": "Zampieri", "suffix": "" }, { "first": "Shervin", "middle": [], "last": "Malmasi", "suffix": "" }, { "first": "Preslav", "middle": [], "last": "Nakov", "suffix": "" }, { "first": "Sara", "middle": [], "last": "Rosenthal", "suffix": "" }, { "first": "Noura", "middle": [], "last": "Farra", "suffix": "" }, { "first": "Ritesh", "middle": [], "last": "Kumar", "suffix": "" } ], "year": 2019, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": { "arXiv": [ "arXiv:1903.08983" ] }, "num": null, "urls": [], "raw_text": "Marcos Zampieri, Shervin Malmasi, Preslav Nakov, Sara Rosenthal, Noura Farra, and Ritesh Kumar. 2019. Semeval-2019 task 6: Identifying and catego- rizing offensive language in social media (offenseval). arXiv preprint arXiv:1903.08983.", "links": null }, "BIBREF56": { "ref_id": "b56", "title": "Character-level Convolutional Networks for Text Classification", "authors": [ { "first": "Xiang", "middle": [], "last": "Zhang", "suffix": "" }, { "first": "Junbo", "middle": [], "last": "Zhao", "suffix": "" }, { "first": "Yann", "middle": [], "last": "Lecun", "suffix": "" } ], "year": 2015, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": { "arXiv": [ "arXiv:1509.01626" ] }, "num": null, "urls": [], "raw_text": "Xiang Zhang, Junbo Zhao, and Yann LeCun. 2015. Character-level Convolutional Networks for Text Classification. arXiv:1509.01626 [cs].", "links": null }, "BIBREF61": { "ref_id": "b61", "title": "Comp/Suff Loss UNIREX (SLM-FP, Comp) 93.59", "authors": [], "year": null, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Comp/Suff Loss UNIREX (SLM-FP, Comp) 93.59 (\u00b10.11) 0.040 (\u00b10.096) 0.350 (\u00b10.048) 0.310 (\u00b10.049) 82.79 (\u00b10.62) 70.74 (\u00b10.81)", "links": null }, "BIBREF63": { "ref_id": "b63", "title": "Suff Criterion UNIREX", "authors": [ { "first": "", "middle": [], "last": "Slm-Fp", "suffix": "" }, { "first": "", "middle": [], "last": "Div", "suffix": "" } ], "year": null, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Suff Criterion UNIREX (SLM-FP, KL Div) 93.06 (\u00b10.25) 0.174 (\u00b10.100) 0.306 (\u00b10.098) 0.131 (\u00b10.005) 82.62 (\u00b10.88) 70.43 (\u00b10.65)", "links": null }, "BIBREF66": { "ref_id": "b66", "title": "SLM Ext Head UNIREX (SLM-FP, Linear) 93", "authors": [], "year": null, "venue": "", "volume": "68", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "SLM Ext Head UNIREX (SLM-FP, Linear) 93.68 (\u00b10.67) 0.189 (\u00b10.030) 0.302 (\u00b10.039) 0.113 (\u00b10.013) 82.55 (\u00b10.84) 70.65 (\u00b10.44)", "links": null }, "BIBREF69": { "ref_id": "b69", "title": "Hate Speech Detection Stormfront Vanilla 10", "authors": [], "year": null, "venue": "", "volume": "48", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Hate Speech Detection Stormfront Vanilla 10.48 (\u00b11.66) -0.066 (\u00b10.072) 0.153 (\u00b10.002) 0.219 (\u00b10.071)", "links": null }, "BIBREF70": { "ref_id": "b70", "title": "Offensive Speech Detection OffenseEval Vanilla", "authors": [], "year": null, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Offensive Speech Detection OffenseEval Vanilla 33.51 (\u00b10.99) -0.125 (\u00b10.068) 0.104 (\u00b10.007) 0.229 (\u00b10.064)", "links": null }, "BIBREF71": { "ref_id": "b71", "title": ") Irony Detection SemEval2018-Irony Vanilla 29", "authors": [], "year": null, "venue": "", "volume": "63", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "UNIREX (DLM-FP) 35.52 (\u00b11.26) 0.053 (\u00b10.012) 0.140 (\u00b10.049) 0.087 (\u00b10.045) UNIREX (SLM-FP) 38.17 (\u00b10.96) 0.039 (\u00b10.031) 0.087 (\u00b10.016) 0.048 (\u00b10.024) Irony Detection SemEval2018-Irony Vanilla 29.63 (\u00b14.72) -0.058 (\u00b10.075) 0.154 (\u00b10.001) 0.212 (\u00b10.074)", "links": null } }, "ref_entries": { "FIGREF0": { "type_str": "figure", "uris": null, "text": "Desiderata of Rationale Extraction. Unlike prior works, UNIREX enables optimizing for all three desiderata.", "num": null }, "FIGREF1": { "type_str": "figure", "uris": null, "text": "UNIREX Framework. UNIREX enables jointly optimizing the task model (Ftask) and rationale extractor (Fext), w.r.t. faithfulness (Lfaith), plausibility (Lplaus), and task performance (Ltask).", "num": null }, "FIGREF2": { "type_str": "figure", "uris": null, "text": "Rationale Extractor Types.", "num": null }, "FIGREF3": { "type_str": "figure", "uris": null, "text": "NRG Comparison by Desiderata. We show FNRG, PNRG, and TNRG for all methods, averaged over all datasets.", "num": null }, "FIGREF4": { "type_str": "figure", "uris": null, "text": "Gold Rationale Data Efficiency on SST.", "num": null }, "FIGREF5": { "type_str": "figure", "uris": null, "text": "Gold Rationale Data Efficiency on CoS-E.", "num": null }, "TABREF0": { "type_str": "table", "html": null, "text": "Comp+Suff) 0.302 (\u00b10.039) 0.113 (\u00b10.013) 82.55 (\u00b10.84) 93.68 (\u00b10.67) FP (MLP-2048-2) 0.323 (\u00b10.071) 0.144 (\u00b10.012) 83.82 (\u00b10.77) 93.67 (\u00b10.18) SLM-FP (MLP-4096-3) 0.295 (\u00b10.057) 0.154 (\u00b10.027) 84.53 (\u00b10.61) 93.19 (\u00b10.79)", "num": null, "content": "
AblationUNIREX ConfigFaithfulness Comp (\u2191) Suff (\u2193)Plausibility AUPRC (\u2191)Performance Acc (\u2191)
Ext Type (F)AA-F (Rand) AA-F (Gold) AA-F (Inv)0.171 (\u00b10.040) 0.327 (\u00b10.050) 44.92 (\u00b10.00) 94.05 (\u00b10.35) 0.232 (\u00b10.088) 0.249 (\u00b10.021) 100.00 (\u00b10.00) 93.81 (\u00b10.54) 0.242 (\u00b10.010) 0.357 (\u00b10.019) 20.49 (\u00b10.00) 93.47 (\u00b11.81)
AA-F (IG)0.292 (\u00b10.051) 0.171 (\u00b10.038) 48.13 (\u00b11.14) 92.97 (\u00b10.44)
AA-FP (Sum)0.296 (\u00b10.067) 0.185 (\u00b10.048) 47.60 (\u00b12.44) 93.25 (\u00b10.45)
Ext Type (FP)AA-FP (MLP) DLM-FP SLM-FP0.285 (\u00b10.051) 0.197 (\u00b10.100) 54.82 (\u00b11.97) 93.23 (\u00b10.92) 0.319 (\u00b10.090) 0.167 (\u00b10.036) 85.80 (\u00b10.74) 93.81 (\u00b10.18) 0.302 (\u00b10.039) 0.113 (\u00b10.013) 82.55 (\u00b10.84) 93.68 (\u00b10.67)
Comp/Suff Loss
SLM-FP (KL Div)
Suff Criterion
" }, "TABREF1": { "type_str": "table", "html": null, "text": "UNIREX Ablation Studies on SST.", "num": null, "content": "" }, "TABREF2": { "type_str": "table", "html": null, "text": "\u00b10.009) 0.258 (\u00b10.031) 93.81 (\u00b10.55) UNIREX (AA-F) 0.292 (\u00b10.051) 0.171 (\u00b10.038) 92.97 (\u00b10.44) UNIREX (DLM-FP) 0.319 (\u00b10.090) 0.167 (\u00b10.036) 93.81 (\u00b10.54) Yelp AA (IG) 0.069 (\u00b10.004) 0.219 (\u00b10.028) 92.50 (\u00b12.07) UNIREX (AA-F) 0.138 (\u00b10.078) 0.126 (\u00b10.059) 83.93 (\u00b113.20) UNIREX (DLM-FP) 0.265 (\u00b10.094) 0.097 (\u00b10.033) 92.37 (\u00b10.46)", "num": null, "content": "
TaskDatasetMethodFaithfulness Comp (\u2191) Suff (\u2193)Task Performance Perf (\u2191)
SASST 0.119 (Amazon AA (IG) AA (IG) 0.076 (\u00b10.010) 0.224 (\u00b10.037) UNIREX (AA-F) 0.130 (\u00b10.077) 0.073 (\u00b10.039) UNIREX (DLM-FP) 0.232 (\u00b10.072) 0.098 (\u00b10.033)91.13 (\u00b10.28) 77.90 (\u00b113.12) 89.35 (\u00b12.22)
HSDStormfrontAA (IG) UNIREX (AA-F) UNIREX (DLM-FP) 0.167 (\u00b10.084) 0.115 (\u00b10.059) 0.135 (\u00b10.010) 0.245 (\u00b10.059) 0.219 (\u00b10.009) 0.092 (\u00b10.025)10.48 (\u00b11.66) 10.36 (\u00b11.94) 10.37 (\u00b12.66)
AA (IG)0.097 (\u00b10.009) 0.244 (\u00b10.052)33.51 (\u00b10.99)
OSD OffenseEvalUNIREX (AA-F)0.074 (\u00b10.040) 0.102 (\u00b10.024)32.62 (\u00b14.85)
UNIREX (DLM-FP) 0.140 (\u00b10.049) 0.087 (\u00b10.045)35.52 (\u00b11.26)
AA (IG)0.128 (\u00b10.014) 0.248 (\u00b10.064)29.63 (\u00b14.72)
IDSemEval2018UNIREX (AA-F) UNIREX (DLM-FP) 0.149 (\u00b10.052) 0.102 (\u00b10.053) 0.069 (\u00b10.041) 0.096 (\u00b10.011)49.95 (\u00b18.31) 31.97 (\u00b12.80)
" }, "TABREF3": { "type_str": "table", "html": null, "text": "Zero-Shot Faithfulness Transfer from SST.", "num": null, "content": "" }, "TABREF5": { "type_str": "table", "html": null, "text": "Plausibility User Study on SST.", "num": null, "content": "
" }, "TABREF7": { "type_str": "table", "html": null, "text": "Main Results on SST.", "num": null, "content": "
MethodComposite NRG (\u2191)NRG (\u2191)Faithfulness Comp (\u2191)Suff (\u2193)NRG (\u2191)Plausibility AUPRC (\u2191)TF1 (\u2191)Performance NRG (\u2191) F1 (\u2191)
AA (Grad)0.4810.4570.184 (\u00b10.023) 0.107 (\u00b10.017)0.02813.31 (\u00b10.91)5.02 (\u00b10.00)0.95795.33 (\u00b10.65)
AA (Input*Grad)0.5030.3590.148 (\u00b10.031) 0.137 (\u00b10.019)0.1948.68 (\u00b10.37)37.58 (\u00b10.55)0.95795.33 (\u00b10.65)
AA (DeepLIFT)0.4680.2590.122 (\u00b10.029) 0.172 (\u00b10.022)0.1879.00 (\u00b10.16)36.15 (\u00b11.45)0.95795.33 (\u00b10.65)
AA (IG)0.4390.1730.134 (\u00b10.016) 0.219 (\u00b10.044)0.1888.88 (\u00b10.21)36.39 (\u00b11.29)0.95795.33 (\u00b10.65)
L2E0.5500.4450.000 (\u00b10.007) 0.026 (\u00b10.015)0.24816.68 (\u00b110.20) 38.92 (\u00b14.07)0.95795.33 (\u00b10.65)
SGT0.5530.4740.124 (\u00b10.053) 0.071 (\u00b10.064)0.18410.05 (\u00b11.23)34.64 (\u00b11.67)1.00096.33 (\u00b10.76)
FRESH0.6450.7320.234 (\u00b10.034) 0.000 (\u00b10.000)0.30517.02 (\u00b16.22)48.26 (\u00b15.87)0.89994.00 (\u00b11.44)
A2R0.4310.7640.267 (\u00b10.050) 0.000 (\u00b10.000)0.24435.44 (\u00b121.69) 19.78 (\u00b125.56)0.28479.78 (\u00b17.14)
UNIREX (AA-F)0.6010.7440.505 (\u00b10.134) 0.122 (\u00b10.100)0.1899.14 (\u00b12.51)36.28 (\u00b11.84)0.87093.33 (\u00b11.61)
SGT+P0.5860.6040.152 (\u00b10.013) 0.022 (\u00b10.004)0.1839.16 (\u00b11.59)35.33 (\u00b10.41)0.97195.66 (\u00b11.16)
FRESH+P0.4910.6910.193 (\u00b10.062) 0.000 (\u00b10.000)0.71065.78 (\u00b111.16) 68.70 (\u00b115.78)0.07074.84 (\u00b112.22)
A2R+P0.5850.7640.267 (\u00b10.076) 0.000 (\u00b10.000)0.99193.53 (\u00b10.93)88.77 (\u00b11.22)0.00073.22 (\u00b10.75)
UNIREX (DLM-P)0.6670.0240.024 (\u00b10.003) 0.238 (\u00b10.004)1.00094.32 (\u00b10.12)89.53 (\u00b11.63)0.97895.83 (\u00b10.29)
UNIREX (AA-FP)0.5430.5140.428 (\u00b10.174) 0.195 (\u00b10.105)0.1938.53 (\u00b10.46)37.71 (\u00b13.12)0.92194.50 (\u00b11.00)
UNIREX (DLM-FP)0.7440.3260.283 (\u00b10.217) 0.216 (\u00b10.005)0.99193.65 (\u00b10.36)88.68 (\u00b12.29)0.91394.33 (\u00b11.61)
UNIREX (SLM-FP)0.7540.3620.313 (\u00b10.059) 0.213 (\u00b10.014)0.96591.70 (\u00b11.84)86.17 (\u00b11.20)0.93594.83 (\u00b10.76)
" }, "TABREF8": { "type_str": "table", "html": null, "text": "Main Results on Movies.", "num": null, "content": "
MethodComposite NRG (\u2191)NRG (\u2191)Faithfulness Comp (\u2191)Suff (\u2193)NRG (\u2191)Plausibility AUPRC (\u2191)TF1 (\u2191)Performance NRG (\u2191) Acc (\u2191)
AA (Grad)0.5370.5040.331 (\u00b10.012) 0.352 (\u00b10.007)0.13037.33 (\u00b10.62) 22.65 (\u00b10.00)0.97763.56 (\u00b11.27)
AA (Input*Grad)0.5730.3610.249 (\u00b10.018) 0.385 (\u00b10.008)0.38339.56 (\u00b10.54) 44.43 (\u00b10.40)0.97763.56 (\u00b11.27)
AA (DeepLIFT)0.6050.3460.254 (\u00b10.035) 0.403 (\u00b10.042)0.49142.82 (\u00b11.83) 51.72 (\u00b11.26)0.97763.56 (\u00b11.27)
AA (IG)0.5780.3270.216 (\u00b10.007) 0.378 (\u00b10.010)0.42940.07 (\u00b15.47) 48.34 (\u00b13.16)0.97763.56 (\u00b11.27)
L2E0.5440.4930.005 (\u00b10.003) 0.010 (\u00b10.008)0.16123.56 (\u00b11.09) 37.80 (\u00b11.10)0.97763.56 (\u00b11.27)
SGT0.6180.3670.197 (\u00b10.040) 0.324 (\u00b10.015)0.49143.68 (\u00b14.68) 51.00 (\u00b13.05)0.99564.35 (\u00b10.46)
FRESH0.3020.5460.037 (\u00b10.036) 0.000 (\u00b10.000)0.26132.35 (\u00b17.66) 39.37 (\u00b10.70)0.10124.81 (\u00b13.46)
A2R0.2770.5160.014 (\u00b10.021) 0.000 (\u00b10.000)0.28241.61 (\u00b13.85) 33.12 (\u00b19.06)0.03221.77 (\u00b11.31)
UNIREX (AA-F)0.6900.5380.297 (\u00b10.141) 0.286 (\u00b10.084)0.55446.97 (\u00b13.41) 53.99 (\u00b11.66)0.97863.58 (\u00b10.61)
SGT+P0.6010.3670.201 (\u00b10.032) 0.328 (\u00b10.022)0.43641.30 (\u00b16.70) 47.95 (\u00b11.65)1.00064.57 (\u00b10.33)
FRESH+P0.3740.5150.013 (\u00b10.021) 0.013 (\u00b10.021)0.60653.40 (\u00b112.87) 53.17 (\u00b17.83)0.00020.36 (\u00b10.66)
A2R+P0.4880.5000.001 (\u00b10.001) 0.000 (\u00b10.000)0.95173.59 (\u00b10.81) 67.63 (\u00b11.54)0.01220.91 (\u00b10.48)
UNIREX (DLM-P)0.7510.2670.180 (\u00b10.016) 0.390 (\u00b10.035)0.99776.07 (\u00b11.63) 69.76 (\u00b10.27)0.99064.13 (\u00b10.46)
UNIREX (AA-FP)0.6850.5510.395 (\u00b10.109) 0.381 (\u00b10.101)0.53745.21 (\u00b14.46) 53.91 (\u00b13.23)0.96863.14 (\u00b10.33)
UNIREX (DLM-FP)0.8140.4920.293 (\u00b10.043) 0.321 (\u00b10.070)0.99776.38 (\u00b10.57) 69.52 (\u00b10.24)0.95362.50 (\u00b11.34)
UNIREX (SLM-FP)0.8070.4940.390 (\u00b10.087) 0.424 (\u00b10.110)0.98375.12 (\u00b10.41) 69.25 (\u00b10.41)0.94462.09 (\u00b12.12)
" }, "TABREF9": { "type_str": "table", "html": null, "text": "Main Results on CoS-E.", "num": null, "content": "
MethodComposite NRG (\u2191)NRG (\u2191)Faithfulness Comp (\u2191)Suff (\u2193)NRG (\u2191)Plausibility AUPRC (\u2191)TF1 (\u2191)Performance NRG (\u2191) F1 (\u2191)
AA (Grad)0.4980.4620.222 (\u00b10.028) 0.120 (\u00b10.018)0.03522.27 (\u00b10.17) 13.81 (\u00b10.00)0.99769.80 (\u00b10.60)
AA (Input*Grad)0.5060.2890.225 (\u00b10.048) 0.260 (\u00b10.059)0.23118.51 (\u00b10.23) 43.45 (\u00b10.05)0.99769.80 (\u00b10.60)
AA (DeepLIFT)0.4930.2490.225 (\u00b10.012) 0.292 (\u00b10.014)0.23418.80 (\u00b10.19) 43.51 (\u00b10.04)0.99769.80 (\u00b10.60)
AA (IG)0.4990.2800.162 (\u00b10.086) 0.222 (\u00b10.086)0.22018.71 (\u00b10.40) 41.79 (\u00b11.33)0.99769.80 (\u00b10.60)
L2E0.5220.3660.007 (\u00b10.006) 0.042 (\u00b10.024)0.20524.48 (\u00b12.71) 32.63 (\u00b16.12)0.99769.80 (\u00b10.60)
SGT0.5940.5640.214 (\u00b10.105) 0.033 (\u00b10.077)0.22418.60 (\u00b10.42) 42.42 (\u00b10.51)0.99569.73 (\u00b10.13)
FRESH0.6750.5710.176 (\u00b10.029) 0.000 (\u00b10.000)0.61724.68 (\u00b17.98) 48.02 (\u00b13.04)0.83864.47 (\u00b13.41)
A2R0.2170.404-0.010 (\u00b10.029) 0.000 (\u00b10.000)0.24918.72 (\u00b10.67) 45.45 (\u00b10.02)0.00036.39 (\u00b10.00)
UNIREX (AA-F)0.7110.9560.505 (\u00b10.050) -0.071 (\u00b10.020)0.23618.82 (\u00b10.40) 43.68 (\u00b10.38)0.93966.17 (\u00b14.58)
SGT+P0.6300.6650.280 (\u00b10.029) 0.283 (\u00b10.039)0.22618.63 (\u00b10.52) 42.71 (\u00b10.39)1.00069.91 (\u00b10.81)
FRESH+P0.4040.4130.000 (\u00b10.013) 0.000 (\u00b10.000)0.73955.87 (\u00b110.13) 63.70 (\u00b19.58)0.06038.41 (\u00b15.34)
A2R+P0.5160.4220.011 (\u00b10.024) 0.000 (\u00b10.000)0.97770.86 (\u00b11.30) 76.21 (\u00b11.68)0.15041.42 (\u00b18.73)
UNIREX (DLM-P)0.7080.1230.127 (\u00b10.010) 0.322 (\u00b10.017)0.99971.80 (\u00b10.27) 77.94 (\u00b10.57)1.00069.91 (\u00b10.76)
UNIREX (AA-FP)0.7061.0000.545 (\u00b10.045) -0.077 (\u00b10.099)0.23119.13 (\u00b10.71) 42.66 (\u00b11.18)0.88866.17 (\u00b14.58)
UNIREX (DLM-FP)0.7510.3270.135 (\u00b10.072) 0.165 (\u00b10.029)0.99871.89 (\u00b10.41) 77.63 (\u00b10.62)0.92967.53 (\u00b11.06)
UNIREX (SLM-FP)0.7840.3770.198 (\u00b10.038) 0.171 (\u00b10.027)0.99771.69 (\u00b10.21) 77.79 (\u00b10.09)0.97969.20 (\u00b11.58)
" }, "TABREF10": { "type_str": "table", "html": null, "text": "Main Results on MultiRC.", "num": null, "content": "
MethodComposite NRG (\u2191)NRG (\u2191)Faithfulness Comp (\u2191)Suff (\u2193)NRG (\u2191)Plausibility AUPRC (\u2191)TF1 (\u2191)Performance NRG (\u2191) F1 (\u2191)
AA (Grad)0.5870.5180.313 (\u00b10.009) 0.380 (\u00b10.025)0.24459.80 (\u00b11.32)15.27 (\u00b10.00)0.99990.78 (\u00b10.27)
AA (Input*Grad)0.5030.2870.205 (\u00b10.005) 0.446 (\u00b10.020)0.22332.98 (\u00b11.37)43.13 (\u00b10.86)0.99990.78 (\u00b10.27)
AA (DeepLIFT)0.5080.2700.195 (\u00b10.012) 0.448 (\u00b10.014)0.25433.47 (\u00b11.31)46.44 (\u00b10.04)0.99990.78 (\u00b10.27)
AA (IG)0.5960.4730.308 (\u00b10.011) 0.414 (\u00b10.020)0.31747.83 (\u00b11.04)37.87 (\u00b11.39)0.99990.78 (\u00b10.27)
L2E0.6060.4600.009 (\u00b10.015) 0.036 (\u00b10.022)0.35858.11 (\u00b10.97)31.35 (\u00b10.27)0.99990.78 (\u00b10.27)
SGT0.5950.5030.288 (\u00b10.025) 0.361 (\u00b10.038)0.29842.46 (\u00b13.03)41.70 (\u00b11.78)0.98590.23 (\u00b10.16)
FRESH0.5180.6610.120 (\u00b10.075) 0.000 (\u00b10.000)0.36138.77 (\u00b16.82)53.71 (\u00b13.30)0.53072.92 (\u00b18.71)
A2R0.2730.5640.053 (\u00b10.048) 0.000 (\u00b10.000)0.25648.48 (\u00b111.14) 29.54 (\u00b124.72)0.00052.72 (\u00b114.08)
UNIREX (AA-F)0.6220.5390.330 (\u00b10.018) 0.383 (\u00b10.055)0.34045.29 (\u00b13.02)43.69 (\u00b11.98)0.98790.31 (\u00b10.19)
SGT+P0.6080.5240.286 (\u00b10.034) 0.339 (\u00b10.032)0.31143.03 (\u00b11.69)42.59 (\u00b11.63)0.98890.36 (\u00b10.08)
FRESH+P0.6140.6950.143 (\u00b10.072) 0.000 (\u00b10.000)0.60356.21 (\u00b110.47) 64.09 (\u00b15.59)0.54473.44 (\u00b112.88)
A2R+P0.8000.7510.182 (\u00b10.097) 0.000 (\u00b10.000)0.99287.30 (\u00b10.44)77.31 (\u00b10.72)0.65677.31 (\u00b10.72)
UNIREX (DLM-P)0.8420.5250.311 (\u00b10.011) 0.371 (\u00b10.032)1.00087.85 (\u00b10.13)77.63 (\u00b10.35)1.00090.80 (\u00b10.33)
UNIREX (AA-FP)0.6260.5290.341 (\u00b10.008) 0.406 (\u00b10.046)0.36344.79 (\u00b10.81)47.18 (\u00b10.83)0.98590.21 (\u00b10.08)
UNIREX (DLM-FP)0.8570.5880.335 (\u00b10.018) 0.346 (\u00b10.023)0.99186.99 (\u00b10.40)77.53 (\u00b10.15)0.99290.51 (\u00b10.12)
UNIREX (SLM-FP)0.8640.6030.353 (\u00b10.017) 0.356 (\u00b10.015)0.99487.58 (\u00b10.14)77.22 (\u00b10.28)0.99490.59 (\u00b10.09)
" }, "TABREF11": { "type_str": "table", "html": null, "text": "Main Results on e-SNLI.", "num": null, "content": "
AblationMethodPerformance Acc (\u2191)CSD (\u2191)Faithfulness Comp (\u2191)Suff (\u2193)Plausibility AUPRC (\u2191) TF1 (\u2191)
UNIREX (AA-F, Rand)
Ext Type (F)
" } } } }