Research directions

How can we tell why a NLP system produces a certain output?

placeholder image 1

How do we know and make sure that a NLP system will work well in the real world?

placeholder image 2

The current SOTA models are huge. How do we make them more efficient?

Explainable and transparent NLP

The NLP systems based on large language models are increasingly deployed in many real-world applications, and have real impact on the lives on many people. However, we still do not have reliable methods to explain the ‘reasoning’ backing their outputs, and many current systems also lack even the minimal transparency about their design and training data.

My current work in this area is backed by 2024 Villum Young Investigator grant (see upcoming positions). It focuses on the problem of attribution of the output of generative language models to their training data. This project has planned collaborations with :us: Allen Institute for AI, :us: Carnegie Mellon University, :us: :fr: HuggingFace, and :jp: RIKEN-CSS.

Some relevant past work:

  1. :newspaper: :large_orange_diamond: Piktus, A., Akiki, C., Villegas, P., Laurençon, H., Dupont, G., Luccioni, S., … Rogers, A. (2023). The ROOTS Search Tool: Data Transparency for LLMs. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations), 304–314. Toronto, Canada: Association for Computational Linguistics.
    ROOTS is a 1.6TB multilingual text corpus developed for the training of BLOOM, currently the largest language model explicitly accompanied by commensurate data governance efforts. In continuation of these efforts, we present the ROOTS Search Tool: a search engine over the entire ROOTS corpus offering both fuzzy and exact search capabilities. ROOTS is the largest corpus to date that can be investigated this way. The ROOTS Search Tool is open-sourced and available on Hugging Face Spaces: We describe our implementation and the possible use cases of our tool.
  2. :newspaper: :large_orange_diamond: Jernite, Y., Nguyen, H., Biderman, S., Rogers, A., Masoud, M., Danchev, V., … Mitchell, M. (2022). Data Governance in the Age of Large-Scale Data-Driven Language Technology. 2022 ACM Conference on Fairness, Accountability, and Transparency, 2206–2222.
    The recent emergence and adoption of Machine Learning technology, and specifically of Large Language Models, has drawn attention to the need for systematic and transparent management of language data. This work proposes an approach to global language data governance that attempts to organize data management amongst stakeholders, values, and rights. Our proposal is informed by prior work on distributed governance that accounts for human values and grounded by an international research collaboration that brings together researchers and practitioners from 60 countries. The framework we present is a multi-party international governance structure focused on language data, and incorporating technical and organizational tools needed to support its work.
  3. :newspaper: González, A. V., Rogers, A., & Søgaard, A. (2021). On the Interaction of Belief Bias and Explanations. Findings of ACL-IJCNLP 2021, 2930–2942. Online: ACL.
    A myriad of explainability methods have been proposed in recent years, but there is little consensus on how to evaluate them. While automatic metrics allow for quick benchmarking, it isn’t clear how such metrics reflect human interaction with explanations. Human evaluation is of paramount importance, but previous protocols fail to account for belief biases affecting human performance, which may lead to misleading conclusions. We provide an overview of belief bias, its role in human evaluation, and ideas for NLP practitioners on how to account for it. For two experimental paradigms, we present a case study of gradient-based explainability introducing simple ways to account for humans’ prior beliefs: models of varying quality and adversarial examples. We show that conclusions about the highest performing methods change when introducing such controls, pointing to the importance of accounting for belief bias in evaluation.
  4. :newspaper: Kovaleva, O., Kulshreshtha, S., Rogers, A., & Rumshisky, A. (2021). BERT Busters: Outlier Dimensions That Disrupt Transformers. Findings of ACL-IJCNLP 2021, 3392–3405. Online: ACL.
    Multiple studies have shown that Transformers are remarkably robust to pruning. Contrary to this received wisdom, we demonstrate that pre-trained Transformer encoders are surprisingly fragile to the removal of a very small number of features in the layer outputs (<0.0001% of model weights). In case of BERT and other pre-trained encoder Transformers, the affected component is the scaling factors and biases in the LayerNorm. The outliers are high-magnitude normalization parameters that emerge early in pre-training and show up consistently in the same dimensional position throughout the model. We show that disabling them significantly degrades both the MLM loss and the downstream task performance. This effect is observed across several BERT-family models and other popular pre-trained Transformer architectures, including BART, XLNet and ELECTRA; we also show a similar effect in GPT-2.
  5. :ledger: Rogers, A., Kovaleva, O., & Rumshisky, A. (2020). A Primer in BERTology: What We Know About How BERT Works. Transactions of the Association for Computational Linguistics, 8, 842–866.
    Transformer-based models have pushed state of the art in many areas of NLP, but our understanding of what is behind their success is still limited. This paper is the first survey of over 150 studies of the popular BERT model. We review the current state of knowledge about how BERT works, what kind of information it learns and how it is represented, common modifications to its training objectives and architecture, the overparameterization issue, and approaches to compression. We then outline directions for future research.
  6. :newspaper: :large_orange_diamond: Kovaleva, O., Romanov, A., Rogers, A., & Rumshisky, A. (2019). Revealing the Dark Secrets of BERT. Proceedings of EMNLP-IJCNLP), 4356–4365.
    BERT-based architectures currently give state-of-the-art performance on many NLP tasks, but little is known about the exact mechanisms that contribute to its success. In the current work, we focus on the interpretation of self-attention, which is one of the fundamental underlying components of BERT. Using a subset of GLUE tasks and a set of handcrafted features-of-interest, we propose the methodology and carry out a qualitative and quantitative analysis of the information encoded by the individual BERT’s heads. Our findings suggest that there is a limited set of attention patterns that are repeated across different heads, indicating the overall model overparametrization. While different heads consistently use the same attention patterns, they have varying impact on performance across different tasks. We show that manually disabling attention in certain heads leads to a performance improvement over the regular fine-tuned BERT models.

Safe and Robust NLP

I use “safety” in the engineering sense of the word: the NLP systems should actually do what their developers are promising, the same way as e.g. construction engineers ensure that the bridges they build withstand the target load. This is inextricably linked to the topic of robustness of generalization: the NLP systems are trained on some data, and to perform in the real world they need to generalize to the real-world data.

My current work in this area is backed by 2023 DFF Inge Lehmann grant (see upcoming positions). It focuses on the development of a benchmark that would reward systems for generalizing rather than memorizing their training data. This project has a planned collaboration with :us: New York University.

Some relevant past work:

  1. :black_nib: Luccioni, A. S., & Rogers, A. (2023). Mind Your Language (Model): Fact-Checking LLMs and Their Role in NLP Research and Practice. arXiv (under review).
    Much of the recent discourse within the NLP research community has been centered around Large Language Models (LLMs), their functionality and potential – yet not only do we not have a working definition of LLMs, but much of this discourse relies on claims and assumptions that are worth re-examining. This position paper contributes a definition of LLMs, explicates some of the assumptions made regarding their functionality, and outlines the existing evidence for and against them. We conclude with suggestions for research directions and their framing in future work.
  2. :ledger: Rogers, A., Gardner, M., & Augenstein, I. (2022). QA Dataset Explosion: A Taxonomy of NLP Resources for Question Answering and Reading Comprehension. ACM CSUR.
    Alongside huge volumes of research on deep learning models in NLP in the recent years, there has been also much work on benchmark datasets needed to track modeling progress. Question answering and reading comprehension have been particularly prolific in this regard, with over 80 new datasets appearing in the past two years. This study is the largest survey of the field to date. We provide an overview of the various formats and domains of the current resources, highlighting the current lacunae for future work. We further discuss the current classifications of “reasoning types" in question answering and propose a new taxonomy. We also discuss the implications of over-focusing on English, and survey the current monolingual resources for other languages and multilingual resources. The study is aimed at both practitioners looking for pointers to the wealth of existing data, and at researchers working on new resources.
  3. :newspaper: :large_orange_diamond: Ray Choudhury, S., Rogers, A., & Augenstein, I. (2022). Machine Reading, Fast and Slow: When Do Models ’Understand’ Language? Proceedings of the 29th International Conference on Computational Linguistics, 78–93. Gyeongju, Republic of Korea.
    Two of the most fundamental issues in Natural Language Understanding (NLU) at present are: (a) how it can established whether deep learning-based models score highly on NLU benchmarks for the ’right’ reasons; and (b) what those reasons would even be. We investigate the behavior of reading comprehension models with respect to two linguistic ’skills’: coreference resolution and comparison. We propose a definition for the reasoning steps expected from a system that would be ’reading slowly’, and compare that with the behavior of five models of the BERT family of various sizes, observed through saliency scores and counterfactual explanations. We find that for comparison (but not coreference) the systems based on larger encoders are more likely to rely on the ’right’ information, but even they struggle with generalization, suggesting that they still learn specific lexical patterns rather than the general principles of comparison.
  4. :newspaper: Bhargava, P., Drozd, A., & Rogers, A. (2021). Generalization in NLI: Ways (Not) To Go Beyond Simple Heuristics. Proceedings of the Second Workshop on Insights from Negative Results in NLP, 125–135. Online and Punta Cana, Dominican Republic: Association for Computational Linguistics.
    Much of recent progress in NLU was shown to be due to models’ learning dataset-specific heuristics. We conduct a case study of generalization in NLI (from MNLI to the adversarially constructed HANS dataset) in a range of BERT-based architectures (adapters, Siamese Transformers, HEX debiasing), as well as with subsampling the data and increasing the model size. We report 2 successful and 3 unsuccessful strategies, all providing insights into how Transformer-based models learn to generalize.
  5. :newspaper: :large_orange_diamond: Rogers, A., Kovaleva, O., Downey, M., & Rumshisky, A. (2020). Getting Closer to AI Complete Question Answering: A Set of Prerequisite Real Tasks. Proceedings of the AAAI Conference on Artificial Intelligence, 11.
    The recent explosion in question answering research produced a wealth of both factoid RC and commonsense reasoning datasets. Combining them presents a different kind of task: not deciding simply whether information is present in the text, but also whether a confident guess could be made for the missing information. To that end, we present QuAIL, the first reading comprehension dataset (a) to combine textbased, world knowledge and unanswerable questions, and (b) to provide annotation that would enable precise diagnostics of the reasoning strategies by a given QA system. QuAIL contains 15K multi-choice questions for 800 texts in 4 domains (fiction, blogs, political news, and user story texts). Crucially, to solve QuAIL a system would need to handle both general and text-specific questions, impossible to answer from pretraining data. We show that the new benchmark poses substantial challenges to the current state-of-the-art systems, with a 30% drop in accuracy compared to the most similar existing dataset.

Safe and Robust NLP

The current cutting-edge NLP systems are based on large language models, some with hundreds with billions of parameters. As they are increasingly embedded in everyday applications and used of millions of people, the carbon costs of their use are also skyrocketing. It is imperative that in the future we find more efficient methods to achieve the same or better levels of performance.

My current work in this area is backed by 2023 DFF Inge Lehmann grant (see upcoming positions). It focuses on the development of a benchmark suite that deliberately caps pre-training and test data, so as to encourage machine learning research on more efficient solutions. This project has a planned collaboration with :us: New York University.

Some relevant past work:

  1. :newspaper: :large_orange_diamond: Prasanna, S., Rogers, A., & Rumshisky, A. (2020). When BERT Plays the Lottery, All Tickets Are Winning. Proceedings of EMNLP, 3208–3229. Online: ACL.
    Much of the recent success in NLP is due to the large Transformer-based models such as BERT (Devlin et al, 2019). However, these models have been shown to be reducible to a smaller number of self-attention heads and layers. We consider this phenomenon from the perspective of the lottery ticket hypothesis. For fine-tuned BERT, we show that (a) it is possible to find a subnetwork of elements that achieves performance comparable with that of the full model, and (b) similarly-sized subnetworks sampled from the rest of the model perform worse. However, the "bad" subnetworks can be fine-tuned separately to achieve only slightly worse performance than the "good" ones, indicating that most weights in the pre-trained BERT are potentially useful. We also show that the "good" subnetworks vary considerably across GLUE tasks, opening up the possibilities to learn what knowledge BERT actually uses at inference time.


Initial recruiting just started! See available and upcoming positions.

Ph.D. and Postdoc Positions

Current funded positions:

Getting your own funding (if you have your own idea for a PhD or postdoc that you’d like to pursue with me):

If that sounds like you, please get in touch, including your CV and transcripts! Also check my upcoming talks, maybe there’s an opportunity to meet at an upcoming conference.


  • Generally, to be enrolled in the Ph.D. school in Denmark, you need to have a Master’s degree.
  • The Danish visa process for non-EU citizens usually takes about 3 months (after you receive and accept the offer).
  • You don’t have to learn Danish.
  • What would it be like to do a PhD in Denmark and ITU? Here’s some tips and reflections by NLPNorth PhD students

Thesis and project supervision at ITU

If you’re a B.Sc. or M.Sc. student at IT University of Copenhagen, and you would like to work with me, please:

  • reach out stating [ITU M.Sc. thesis], [ITU B.Sc. thesis] or [ITU research project] at the start of the subject.
  • introduce yourself and state:
    • a few concrete topics that you are interested in,
    • your background and any relevant experience.

Here are some of the research directions that I would be interested in:

  • Language model analysis: identifying the types of knowledge acquired from language model pre-training
  • Language processing strategies: do NLP models perform well for the right reasons? What strategies should they follow when solving reasoning tasks?
  • Robustness and generalization: do NLP models reliably perform their tasks out of training distribution, and what can we do to help them?
  • NLP system auditing and documentation: establishing the cases where a system is safe to deploy
  • NLP system interpretability: how can we establish how a deep learning model arrives at its decisions?
  • Sustainable NLP: how can we build systems that work well, but don’t require billions of parameters and terabytes of data?

Feel free to also look at my publications and see if there’s anything you’d like to build on.