Papers by Yanai Elazar
First Align, then Predict: Understanding the Cross-Lingual Ability of Multilingual BERT (2021.eacl-main)
Copied to clipboard
| Challenge: | Multilingual pretrained language models have demonstrated remarkable zero-shot cross-lingual transfer capabilities. |
| Approach: | They propose to use a layer ablation technique to create a multilingual model that is viewed as a stacking of two sub-networks: a language-agnostic encoder and a task-specific predictor. |
| Outcome: | The proposed model can perform zero-shot cross-lingual transfer for many languages. |
Detection and Measurement of Syntactic Templates in Generated Text (2024.emnlp-main)
Copied to clipboard
| Challenge: | Existing diversity evaluation focuses primarily on word-level features. |
| Approach: | They propose a method for evaluating diversity over syntactic features to characterize general repetition in large language models. |
| Outcome: | The proposed method shows that models produce templated text in downstream tasks at a higher rate than what is found in human-reference texts. |
Contrastive Explanations for Model Interpretability (2021.emnlp-main)
Copied to clipboard
| Challenge: | Existing methods for producing model explanations seek all causal factors at once, making them difficult to comprehend. |
| Approach: | They propose a method to produce contrastive explanations in the latent space . they use attribution and token/span attribution to produce models that consider only contrastive reasoning . |
| Outcome: | The proposed method allows model behavior to consider only contrastive reasoning . it also uncovers which aspects of the input are useful for and against particular decisions . |
CIKQA: Learning Commonsense Inference with a Unified Knowledge-in-the-loop QA Paradigm (2023.findings-eacl)
Copied to clipboard
| Challenge: | Existing commonsense reasoning datasets target different knowledge types, modalities, and formats, but how to help machines acquire and infer over commonsensical knowledge is still unclear. |
| Approach: | They propose a commonsense reasoning benchmark to motivate commonsensing progress from two perspectives: (1) Evaluating whether models can distinguish knowledge quality by predicting if the knowledge is enough to answer the question or not. |
| Outcome: | The proposed model outperforms existing models in evaluating their generalization capabilities across tasks while demonstrating that distinguishing knowledge quality remains challenging for current models. |
Adversarial Removal of Demographic Attributes Revisited (D19-1)
Copied to clipboard
| Challenge: | Several approaches have been proposed to learn classifiers that are invariant (unbiased with respect) to protected attributes. |
| Approach: | They propose to use a diagnostic classifier trained on a held-out subsample to find protected attributes for mention detection at above-chance levels. |
| Outcome: | The proposed classifier generalizes poorly to new in-domain and new domains, suggesting it relies on correlations specific to their particular data sample. |
oLMpics-On What Language Model Pre-training Captures (2020.tacl-1)
Copied to clipboard
| Challenge: | Recent success of pre-trained language models has spurred widespread interest in their capabilities. |
| Approach: | They propose an evaluation protocol that includes zero-shot evaluation and no fine-tuning . they propose to compare the learning curve of a fine- tuned LM to the learning of multiple controls . |
| Outcome: | The proposed evaluation protocol compares the learning curve of a fine-tuned LM to the learning of multiple controls. |
OLMoTrace: Tracing Language Model Outputs Back to Trillions of Training Tokens (2025.acl-demo)
Copied to clipboard
Jiacheng Liu, Taylor Blanton, Yanai Elazar, Sewon Min, Yen-Sung Chen, Arnavi Chheda-Kothary, Huy Tran, Byron Bischoff, Eric Marsh, Michael Schmitz, Cassidy Trier, Aaron Sarnat, Jenna James, Jon Borchardt, Bailey Kuehl, Evie Yu-Yen Cheng, Karen Farley, Taira Anderson, David Albright, Carissa Schoenick, Luca Soldaini, Dirk Groeneveld, Rock Yuren Pang, Pang Wei Koh, Noah A. Smith, Sophie Lebrecht, Yejin Choi, Hannaneh Hajishirzi, Ali Farhadi, Jesse Dodge
| Challenge: | tracing language models' outputs back to training data is a problem because they are trained on text corpora with trillions of tokens . existing methods for tracers have not been scaled to work within this multi-trillion-token setting . |
| Approach: | They propose a system that traces language models' outputs verbatim back to training data . OLMOTRACE retrieves documents from the model's training data that contain exact matches . |
| Outcome: | The proposed system can find verbatim matches between LM output and training data . it can be used to explore fact checking, hallucination, and creativity of language models . |
Revisiting Few-shot Relation Classification: Evaluation Data and Classification Schemes (2021.tacl-1)
Copied to clipboard
| Challenge: | a recent study has focused on few-shot learning (FSL) for relation classification, but it requires large amounts of training data. |
| Approach: | They propose a method for deriving more realistic few-shot test data from available datasets for supervised RC. |
| Outcome: | The proposed method yields a challenging benchmark for FSL-RC on which state of the art models show poor performance. |
Measuring and Improving Consistency in Pretrained Language Models (2021.tacl-1)
Copied to clipboard
Yanai Elazar, Nora Kassner, Shauli Ravfogel, Abhilasha Ravichander, Eduard Hovy, Hinrich Schütze, Yoav Goldberg
| Challenge: | In this paper, we examine whether pretrained language models are consistent with factual knowledge. |
| Approach: | They propose a method to improve consistency of pretrained language models . consistency is a desirable property of a good language understanding model, they argue . |
| Outcome: | The proposed model improves consistency and shows that it is effective. |
Measuring and Improving Attentiveness to Partial Inputs with Counterfactuals (2024.findings-emnlp)
Copied to clipboard
Yanai Elazar, Bhargavi Paranjape, Hao Peng, Sarah Wiegreffe, Khyathi Chandu, Vivek Srikumar, Sameer Singh, Noah Smith
| Challenge: | Existing studies have found that datasets with paired inputs are prone to spurious correlations, resulting in models trained only on those outperform chance. |
| Approach: | They propose a counterfactual attentiveness test to measure reliance on spurious correlations by replacing part of the input with its counterpart from a different example. |
| Outcome: | The proposed method improves models' attentiveness on ten datasets spanning four tasks: natural language inference, reading comprehension, paraphrase detection, and visual & language reasoning. |
Evaluating n-Gram Novelty of Language Models Using Rusty-DAWG (2024.emnlp-main)
Copied to clipboard
| Challenge: | a new study examines how novel language models generate training text . large LMs and constrained decoding strategies both decrease novelty . |
| Approach: | They develop a novel search tool inspired by genomic data to find n-grams in training data. |
| Outcome: | The proposed tool can search for n-grams over a corpus in constant time w.r.t. large LMs and more constrained decoding strategies both decrease novelty. |
Erratum: Measuring and Improving Consistency in Pretrained Language Models (2021.tacl-1)
Copied to clipboard
Yanai Elazar, Nora Kassner, Shauli Ravfogel, Abhilasha Ravichander, Eduard Hovy, Hinrich Schütze, Yoav Goldberg
| Challenge: | During production of this paper, an error was introduced to the formula on the bottom of the right column of page 1020. |
| Approach: | the formula was changed in the last two terms of the paper . |
| Outcome: | the correct formula is now available on the web. |
Amnesic Probing: Behavioral Explanation with Amnesic Counterfactuals (2021.tacl-1)
Copied to clipboard
| Challenge: | Amnesic probing is a method that focuses on how information is being used, rather than on what information is encoded. |
| Approach: | They propose a method that focuses on how the information is being used rather than on what information is encoded. |
| Outcome: | The proposed method is based on a BERT dataset to ask questions that were not possible before . it shows that probing performance is not correlated to task importance . |
Null It Out: Guarding Protected Attributes by Iterative Nullspace Projection (2020.acl-main)
Copied to clipboard
| Challenge: | Word embeddings, pre-trained language models, and deep learning methods are becoming effective for text classification. |
| Approach: | They propose a method for removing information from neural representations using null-space projection. |
| Outcome: | The proposed method mitigates bias in word embeddings and increases fairness in multi-class classification. |
Adversarial Removal of Demographic Attributes from Text Data (D18-1)
Copied to clipboard
| Challenge: | Recent advances in Representation Learning and Adversarial Training remove unwanted features from the learned representation. |
| Approach: | They show that demographic information of authors is encoded in the intermediate representations learned by text-based neural classifiers. |
| Outcome: | The proposed approach achieves higher accuracies on the same dataset, the authors show . they show that the proposed approach is effective in removing unwanted features from the learned representations. |
Text-based NP Enrichment (2022.tacl-1)
Copied to clipboard
| Challenge: | Existing NLP tasks and benchmarks do not cover all NP-mediated relations . we aim to enrich each NP in a text with all the preposition-mediated relationships that hold between it and other NPs in the text. |
| Approach: | They propose a task to enrich NPs with preposition-mediated relations that hold between them . they build a large-scale dataset and analyze the data to test the task . |
| Outcome: | The proposed task is based on a large-scale dataset and fine-tuned language models. |
Lexical Generalization Improves with Larger Models and Longer Training (2022.findings-emnlp)
Copied to clipboard
| Challenge: | Excessive utilization of lexical overlap heuristics can lead to failure on challenging inputs. |
| Approach: | They analyze the use of lexical overlap heuristics in natural language inference, paraphrase detection, and reading comprehension using a contrastive dataset. |
| Outcome: | The proposed model can be used to improve performance on a wide range of tasks, but it is often wrong. |
Evaluating Models’ Local Decision Boundaries via Contrast Sets (2020.findings-emnlp)
Copied to clipboard
Matt Gardner, Yoav Artzi, Victoria Basmov, Jonathan Berant, Ben Bogin, Sihao Chen, Pradeep Dasigi, Dheeru Dua, Yanai Elazar, Ananth Gottumukkala, Nitish Gupta, Hannaneh Hajishirzi, Gabriel Ilharco, Daniel Khashabi, Kevin Lin, Jiangming Liu, Nelson F. Liu, Phoebe Mulcaire, Qiang Ning, Sameer Singh, Noah A. Smith, Sanjay Subramanian, Reut Tsarfaty, Eric Wallace, Ally Zhang, Ben Zhou
| Challenge: | Standard test sets for supervised learning evaluate in-distribution generalization but are misleading when a dataset has systematic gaps. |
| Approach: | They propose a more rigorous annotation paradigm for NLP that helps to close systematic gaps in the test data. |
| Outcome: | The proposed model performs significantly lower on contrast sets than on the original test sets—up to 25% in some cases. |
Hybrid Preferences: Learning to Route Instances for Human vs. AI Feedback (2025.acl-long)
Copied to clipboard
Lester James Validad Miranda, Yizhong Wang, Yanai Elazar, Sachin Kumar, Valentina Pyatkin, Faeze Brahman, Noah A. Smith, Hannaneh Hajishirzi, Pradeep Dasigi
| Challenge: | Learning from human feedback has enabled the alignment of language models (LMs) with human preferences. |
| Approach: | They propose a Hybrid Preference routER that defers an annotation to either humans or LMs, achieving better annotation quality while reducing the cost of human-only annotation. |
| Outcome: | The proposed model achieves better annotation quality while reducing the cost of human-only annotation. |
Applying Intrinsic Debiasing on Downstream Tasks: Challenges and Considerations for Machine Translation (2024.emnlp-main)
Copied to clipboard
| Challenge: | In this study, we examine three considerations for intrinsic debiasing in neural machine translation models. |
| Approach: | They propose to measure the extrinsic bias of neural machine translation models by embedding them in a neural embeddable space and using different tokens to debias them. |
| Outcome: | The proposed methods over-rely on gender stereotypes and over-represent them in their models. |
Back to Square One: Artifact Detection, Training and Commonsense Disentanglement in the Winograd Schema (2021.emnlp-main)
Copied to clipboard
| Challenge: | Pre-trained language models have boosted performance on some WS benchmarks, but the source of improvement is not clear. |
| Approach: | They propose a method that uses twin sentences for evaluation and two new baselines that account for artifacts in WS benchmarks. |
| Outcome: | The proposed evaluation method is suboptimal for the Winograd Schema . it uses twin sentences to account for commonsense reasoning abilities . |
Do Language Embeddings capture Scales? (2020.findings-emnlp)
Copied to clipboard
| Challenge: | Pretrained Language Models possess significant linguistic, common sense and factual knowledge, but are short of the capability required for general common-sense reasoning. |
| Approach: | They propose to train pretrained language models with a method of canonicalizing numbers . they address a task which is also pre-requisite for general common-sense reasoning . |
| Outcome: | The proposed model can answer questions about common sense and linguistics, but lacks the capability to answer questions on scalar attributes. |
Few-shot Fine-tuning vs. In-context Learning: A Fair Comparison and Evaluation (2023.findings-acl)
Copied to clipboard
| Challenge: | Recent studies show that in-context learning and few-shot fine-tuning can generalize well out-of-domain. |
| Approach: | They compare few-shot fine-tuning and in-context learning for task adaptation . they find that both approaches generalize similarly, but exhibit large variation . |
| Outcome: | The proposed methods outperform in-context learning and few-shot fine-tuning with OPT models of different sizes. |
How Large Are Lions? Inducing Distributions over Quantitative Attributes (P19-1)
Copied to clipboard
| Challenge: | Current NLP systems have little knowledge about quantitative attributes of objects and events. |
| Approach: | They propose to use web data to create a resource consisting of distributions over physical quantities associated with objects, adjectives, and verbs. |
| Outcome: | The proposed method compares favorably with state-of-the-art results on existing datasets for relative comparisons of nouns and adjectives and on a new dataset. |
The Bias Amplification Paradox in Text-to-Image Generation (2024.naacl-long)
Copied to clipboard
| Challenge: | amplification is a phenomenon in which models exacerbate biases or stereotypes in training data. |
| Approach: | They compare gender ratios in training vs. generated images to investigate bias amplification . they find that a model amplifys gender-occupation biases considerably . |
| Outcome: | The proposed model amplifys gender-occupation biases in training data, but it can be attributed to discrepancies between training captions and model prompts. |