Papers with model’s representations
Leveraging Prototypical Representations for Mitigating Social Bias without Demographic Information (2024.naacl-short)
Copied to clipboard
| Challenge: | Existing approaches to mitigate social biases require explicit annotation of demographic information for each sample. |
| Approach: | They propose a method that leverages predefined demographic texts and incorporates a regularization term during the fine-tuning process to mitigate bias in language models. |
| Outcome: | The proposed method outperforms debiasing methods with limited demographic-annotated data. |
Debiasing Methods in Natural Language Understanding Make Bias More Accessible (2021.emnlp-main)
Copied to clipboard
| Challenge: | Recent debiasing methods in natural language understanding improve performance on out-of-distribution datasets by pressuring models into making unbiased predictions. |
| Approach: | They propose a general probing-based framework that allows for post-hoc interpretation of biases in language models and use an information-theoretic approach to measure the extractability of certain biase . |
| Outcome: | The proposed framework allows for post-hoc interpretation of biases in language models and measures the extractability of certain biase . |
A Practical Method for Generating String Counterfactuals (2025.findings-naacl)
Copied to clipboard
| Challenge: | Interventions targeting the representation space of language models (LMs) have emerged as an effective means to influence model behavior. |
| Approach: | They propose a method to convert representation counterfactuals into string counterf actuals and analyze the linguistic alterations corresponding to the intervention. |
| Outcome: | The proposed method analyzes linguistic alterations corresponding to a given representation space intervention and interprets features utilized to encode a specific concept. |
Personas as a Way to Model Truthfulness in Language Models (2024.emnlp-main)
Copied to clipboard
| Challenge: | Large language models are trained on vast amounts of text from the internet, which contains factual and misleading information. |
| Approach: | They hypothesize that the pretraining data is generated by groups of (un)truthful agents whose outputs share common features and form a (un-truthfully persona) this allows the model to separate truth from falsehoods and controls the truthfulness of its generation. |
| Outcome: | The proposed model can infer truth from falsehoods by finetuning its model on a set of facts and finetuned it on unseen topics. |
Shielded Representations: Protecting Sensitive Attributes Through Iterative Gradient-Based Projection (2023.findings-acl)
Copied to clipboard
| Challenge: | Natural language processing models tend to learn and encode social biases present in the data. |
| Approach: | They propose a method for removing non-linear encoded concepts from neural representations by iteratively training neural classifiers to predict a particular attribute, followed by a projection of the representation on a hypersurface. |
| Outcome: | The proposed method removes non-linear encoded concepts from neural representations. |
Probing for the Usage of Grammatical Number (2022.acl-long)
Copied to clipboard
| Challenge: | Pre-trained language models can be used to perform a wide array of NLP tasks, but their encoding is still a mystery. |
| Approach: | They propose a usage-based probing setup to find an encoding that the model actually uses, and propose 'a use-based approach' they propose to use a behavioral task to remove the linguistic property, and to identify which encodes are used to transfer information from a noun to its head verb. |
| Outcome: | The proposed encodings are based on a behavioral task which cannot be solved without the linguistic property. |