Papers by Kawin Ethayarajh
Rotate King to get Queen: Word Relationships as Orthogonal Transformations in Embedding Space (D19-1)
Copied to clipboard
| Challenge: | Word embeddings have a notable property that word relationships can exist as linear substructures in the embeddable space. |
| Approach: | They propose an alternative way in which downstream models might learn these relationships: orthogonal and linear transformations. |
| Outcome: | The proposed model can learn such relationships as geometric translations, but there is no evidence that it is exclusively accurate. |
Understanding Undesirable Word Embedding Associations (P19-1)
Copied to clipboard
| Challenge: | Word embeddings are often criticized for capturing undesirable word associations such as gender stereotypes. |
| Approach: | They propose to use subspace projection to debias vectors post hoc using a model that implicitly does matrix factorization to debunk gender bias. |
| Outcome: | The proposed test overestimates gender bias in word embeddings by using subspace projection, a method that is widely used in training. |
Attention Flows are Shapley Value Explanations (2021.acl-short)
Copied to clipboard
| Challenge: | Shapley Values are a popular type of explanation in machine learning, but leave-one-out and attention-based explanations still predominate in NLP. |
| Approach: | They propose to use attention flow to explain the importance of features, embeddings, and even neurons to explain credit assignment problems in cooperative game theory. |
| Outcome: | The proposed explanations can explain the importance of features, embeddings, and even neurons, but in NLP, leave-one-out and attention-based explanations still predominate. |
Problems with Cosine as a Measure of Embedding Similarity for High Frequency Words (2022.acl-short)
Copied to clipboard
| Challenge: | We find that word similarities estimated by cosine over contextual embeddings are understated and trace this effect to training data frequency. |
| Approach: | They propose to use cosine similarity to estimate word similarities in contextual embeddings to trace this effect to training data frequency. |
| Outcome: | The proposed model underestimates similarity between frequent and low frequency words even after controlling for polysemy and other factors. |
Utility is in the Eye of the User: A Critique of NLP Leaderboards (2020.emnlp-main)
Copied to clipboard
| Challenge: | Performance-based evaluation has been at the expense of other attributes valued by the NLP community, such as compactness and energy efficiency. |
| Approach: | They propose to frame both the leaderboard and NLP practitioners as consumers and the benefit they get from a model as its utility to them. |
| Outcome: | The proposed model size and energy efficiency benchmarks have been successful in driving the creation of more accurate models, but have been at the expense of other attributes valued by the NLP community. |
Richer Countries and Richer Representations (2022.findings-acl)
Copied to clipboard
| Challenge: | Using BERT, countries with low frequency in training data are less likely to be invocabulary, and are less frequently predicted in the masked language modeling task. |
| Approach: | They propose three criteria to characterize the quality of representations for particular entities or groups: consistency, distinctiveness, and recognizability. |
| Outcome: | The results suggest that frequency is highly correlated with a country’s GDP, perpetuating historic power and wealth inequalities. |
How Contextual are Contextualized Word Representations? Comparing the Geometry of BERT, ELMo, and GPT-2 Embeddings (D19-1)
Copied to clipboard
| Challenge: | Existing word embeddings were static, requiring all senses of a polysemous word to share the same representation. |
| Approach: | They found that the contextualized representations of all words are not isotropic in any layer of the contextualizing model. |
| Outcome: | The results show that the representations of all words are not isotropic in any layer of the contextualizing model. |
Towards Understanding Linear Word Analogies (P19-1)
Copied to clipboard
| Challenge: | Existing theories of word embeddings make strong assumptions about the embeddable space or word distribution. |
| Approach: | They propose a formal explanation of word analogies by adding arithmetic operators to non-linear embedding models such as skip-gram with negative sampling. |
| Outcome: | The proposed model downweights the more frequent word, as weighting schemes do ad hoc. |
The Authenticity Gap in Human Evaluation (2022.emnlp-main)
Copied to clipboard
| Challenge: | Using the standard protocol to evaluate NLGs is often violated, resulting in annotator ratings cease to reflect their preferences. |
| Approach: | They propose a human evaluation protocol called system-level probabilistic assessment (SPA) this protocol is based on the assumption that annotators are biased by likert scales . |
| Outcome: | The proposed protocol can recover the ordering of GPT-3 models by size, but less than half of the expected preferences can be recovered when human evaluation is done with the standard protocol. |
Is Your Classifier Actually Biased? Measuring Fairness under Uncertainty with Bernstein Bounds (2020.acl-main)
Copied to clipboard
| Challenge: | Most NLP datasets are not annotated with protected attributes such as gender, making it difficult to measure classification bias using standard measures of fairness. |
| Approach: | They propose to use Bernstein bounds to represent uncertainty about bias estimate as a confidence interval. |
| Outcome: | The proposed method prevents classifiers from being deemed biased or unbiased when there is insufficient evidence to make either claim. |
Conditional probing: measuring usable information beyond a baseline (2021.emnlp-main)
Copied to clipboard
| Challenge: | Existing methods for probing representations are limited to predicting part-of-speech . current methods cannot detect when a representation is predictive of just aspects of part- of-seech not explainable by the word identity. |
| Approach: | They propose to condition on the information in a baseline representation to test whether it is predictive of part-of-speech. |
| Outcome: | The proposed method is based on a theory of usable information called V-information and conditions on the information in the baseline. |
Anchor Points: Benchmarking Models with Much Fewer Examples (2024.eacl-long)
Copied to clipboard
| Challenge: | Modern language models exhibit powerful but brittle behavior, leading to larger and more diverse benchmarks. |
| Approach: | They propose to use anchor points to select small subsets of a language model-prompt dataset to capture model behavior across the entire dataset. |
| Outcome: | The proposed technique outperforms standard benchmarks in language models with 1-30 anchor points . the proposed technique can be used to compare models on different regions of the dataset . |