Papers by Oana-Maria Camburu
Identifying Linear Relational Concepts in Large Language Models (2024.naacl-long)
Copied to clipboard
| Challenge: | a technique for finding concept directions for human-interpretable concepts is needed to find their direction in the latent space . a linear relational concept (LRC) can be used to locate concepts in hidden activations . |
| Approach: | They propose a method for finding human-interpretable concepts by inverting a linear relational embedding and using earlier object layers. |
| Outcome: | The proposed method outperforms standard probing classifiers on performance as concept classifier and ability to causally change model output. |
Does the Objective Matter? Comparing Training Objectives for Pronoun Resolution (2020.emnlp-main)
Copied to clipboard
| Challenge: | Pre-trained language models have been used to obtain state-of-the-art results on pronoun resolution. |
| Approach: | They propose to use pre-trained language models to train pronoun resolution models . they compare performance and seed-wise stability of four models that represent four objectives . |
| Outcome: | The proposed model performs best in-domain, while the other objectives perform poorly out-of-domain. |
WikiCREM: A Large Unsupervised Corpus for Coreference Resolution (D19-1)
Copied to clipboard
| Challenge: | Large-scale training sets for pronoun resolution are scarce, since manually labelling data is costly. |
| Approach: | They propose a language-model-based approach to solve pronoun disambiguation problems using a WikiCREM dataset. |
| Outcome: | The proposed model outperforms state-of-the-art approaches on 6 out of 7 datasets. |
Using Natural Language Explanations to Improve Robustness of In-context Learning (2024.acl-long)
Copied to clipboard
| Challenge: | Recent studies show that large language models excel in many tasks via in-context learning (ICL). However, ICL struggles to execute complex tasks such as arithmetic, commonsense, and symbolic reasoning. |
| Approach: | They propose to augment ICL with natural language explanations (NLEs) to produce further NLEs on adversarial datasets. |
| Outcome: | The proposed approach yields more accurate results than zero-shot-ICL and using only human-generated NLEs on eight adversarial datasets. |
Few-Shot Out-of-Domain Transfer Learning of Natural Language Explanations in a Label-Abundant Setup (2022.findings-emnlp)
Copied to clipboard
| Challenge: | Existing approaches to train models to provide natural language explanations (NLEs) require acquisition of task-specific NLEs, which is time- and resource-consuming. |
| Approach: | They propose a few-shot out-of-domain transfer of NLEs from a parent task to a child task . they propose four methods that cover possible fine-tuning combinations of NLESs and labels . |
| Outcome: | The proposed methods cover the possible fine-tuning combinations of labels and NLEs for the parent and child tasks. |
KNOW How to Make Up Your Mind! Adversarially Detecting and Alleviating Inconsistencies in Natural Language Explanations (2023.acl-short)
Copied to clipboard
| Challenge: | eIA is an adversarial attack that generates inconsistent natural language explanations (NLEs) a model that generate In-NLE is undesirable, as it has a faulty decision-making process or is prone to inconsistencies. |
| Approach: | They propose an off-the-shelf mitigation method to alleviate inconsistencies by grounding the model into external background knowledge. |
| Outcome: | The proposed method reduces inconsistencies detected by previous models . it is based on external knowledge bases and a novel approach to mitigate inconsistent models based upon the proposed method . |
SparseFit: Few-shot Prompting with Sparse Fine-tuning for Jointly Generating Predictions and Natural Language Explanations (2024.acl-long)
Copied to clipboard
| Challenge: | Models that generate natural language explanations (NLEs) for their predictions often require large datasets of human-written NLEs at training time, which can be expensive and time-consuming to collect. |
| Approach: | They propose a sparse few-shot fine-tuning strategy that leverages discrete prompts to jointly generate predictions and NLEs. |
| Outcome: | The proposed approach compares sparse few-shot fine-tuning with existing parametric fine- tuning techniques on three sizes of the T5 language model and four datasets and produces competitive results for both task performance and NLE quality. |
Fool Me Once? Contrasting Textual and Visual Explanations in a Clinical Decision-Support Setting (2024.emnlp-main)
Copied to clipboard
Maxime Kayser, Bayar Menzat, Cornelius Emde, Bogdan Bercean, Alex Novak, Abdalá Morgado, Bartlomiej Papiez, Susanne Gaube, Thomas Lukasiewicz, Oana-Maria Camburu
| Challenge: | XAI models are being used in safety-critical domains, but their use is limited due to their limited transparency and insufficient model robustness. |
| Approach: | They evaluated visual, natural language and a combination of both modalities to examine how users use them. |
| Outcome: | The proposed model is more robust and transparent than previous models. |
A Surprisingly Robust Trick for the Winograd Schema Challenge (P19-1)
Copied to clipboard
| Challenge: | The Winograd Schema Challenge (WSC) dataset WSC273 and its inference counterpart WNLI are popular benchmarks for natural language understanding and commonsense reasoning. |
| Approach: | They propose to fine-tune language models on the Winograd Schema Challenge dataset WSC273 and its inference counterpart WNLI to achieve accuracies of 72.5% and 74.7%, respectively. |
| Outcome: | The proposed language models achieve 72.5% and 74.7% accuracy on the WSC273 and WNLI datasets, respectively. |
The Probabilities Also Matter: A More Faithful Metric for Faithfulness of Free-Text Explanations in Large Language Models (2024.acl-short)
Copied to clipboard
| Challenge: | In many applications of ML systems it is important to understand why the system came to a particular answer. |
| Approach: | They introduce a faithfulness metric based on counterfactual input edits that takes into account not just the binary label change, but the total shift in the model’s predicted label distribution. |
| Outcome: | The proposed explanations are more likely to mention factors when they are impactful to the model’s prediction, with the degree of association increasing with model size but varying significantly by task. |
Counter-GAP: Counterfactual Bias Evaluation through Gendered Ambiguous Pronouns (2023.eacl-main)
Copied to clipboard
| Challenge: | a number of studies have focused on gender bias in language models, but these methods fail to detect it. |
| Approach: | They propose to use gender bias in coreference resolution to evaluate gender bias . they propose to construct an annotated quadruple-level dataset with 4008 instances . |
| Outcome: | The proposed method is able to detect gender bias in a quadruple dataset . previous methods failed to detect bias or cancel it, the authors argue . |
Atomic Inference for NLI with Generated Facts as Atoms (2024.emnlp-main)
Copied to clipboard
| Challenge: | Existing models that can provide accurate explanations are not interpretable, i.e. they do not reflect the inner workings of the model. |
| Approach: | They propose to use LLM-generated facts as atoms to make interpretable models that can be used to make accurate predictions for each component part of an input. |
| Outcome: | The proposed method outperforms existing methods on natural language understanding tasks with a multi-stage fact generation process and a training regime that incorporates the facts. |
Make Up Your Mind! Adversarial Generation of Inconsistent Natural Language Explanations (2020.acl-main)
Copied to clipboard
| Challenge: | a promising research direction consists of designing neural models capable of generating natural language explanations for their predictions. |
| Approach: | They propose a framework for sanity checking models against inconsistent explanations . they apply the framework to a state-of-the-art neural natural language inference model . |
| Outcome: | The proposed framework can generate inconsistent explanations on a state-of-the-art model . it also addresses the problem of adversarial attacks with full target sequences . |
Faithfulness Tests for Natural Language Explanations (2023.acl-short)
Copied to clipboard
Pepa Atanasova, Oana-Maria Camburu, Christina Lioma, Thomas Lukasiewicz, Jakob Grue Simonsen, Isabelle Augenstein
| Challenge: | Existing methods for explaining neural models are misleading as they often present reasons that are unfaithful to the model’s inner workings. |
| Approach: | They propose a counterfactual input editor for inserting reasons that lead to counterfacts but are not reflected by the NLEs. |
| Outcome: | The proposed model can evaluate emerging NLE models, proving a fundamental tool in the development of faithful explanations. |