Analyzing and interpreting neural networks for NLP

Revealing the content of the neural black box: workshop on the analysis and interpretation of neural networks for Natural Language Processing.

This project is maintained by blackboxnlp


The workshop will be collocated with EMNLP 2020.

Outstandig paper awards

Two BlackboxNLP 2020 papers were selected for the outstanding paper award:


Important dates (updated!)

Workshop programme

BlackboxNLP2020 will be held virtually. To promote accessibility across timezones, the programme is divided into three blocks. All talks and posters will be presented twice, in two different blocks. A detailed version of the programme is available here.

All plenary sessions will be held in zoom, where we will livestream the keynote and oral presentations, that are followed by a live Q&A. Questions can be asked in rocket chat during the presentations. Poster sessions will be held in gather town. The links to the zoom session, space and the rocketchat channel can be found on the EMNLP page of the workshop (only accessible with conference registration).

Block B

Session Location Time      
    US West coast Punta Cana London China
    (UTC-8) (UTC-4) (UTC) (UTC+8)
Opening remarks Zoom 00:00 - 00:15 4:00 - 4:15 8:00 - 8:15 16:00 - 16:15
Keynote speaker 1 – Anna Rogers Zoom 00:15 - 01:00 4:15 - 5:00 8:15 - 9:00 16:15 - 17:00
Oral presentations 1 Zoom 01:15 - 02:00 5:15 - 6:00 9:15 - 10:00 17:15 - 18:00
Demo presentation Zoom 02:15 - 02:30 6:15 - 6:30 10:15 - 10:30 18:15 - 18:30
Poster session B – room K-N 02:30 - 04:00 6:30 - 8:00 10:30 - 12:00 18:30 - 20:00
Keynote speaker 1 – Anna Rogers Zoom 04:15 - 05:00 8:15 - 9:00 12:15 - 13:00 20:15 - 21:00

Block C

Session Location Time      
    US West coast Punta Cana London China
    (UTC-8) (UTC-4) (UTC) (UTC+8)
Keynote speaker 2 – Roger Levy Zoom 7:00 - 7:45 11:00 - 11:45 15:00 - 15:45 19:00 - 19:45
Oral presentations 2 Zoom 8:00 - 9:00 12:00 - 13:00 16:00 - 17:00 20:00 - 21:00
Keynote speaker 3 – Idan Blank Zoom 9:15 - 10:00 13:15 - 14:00 17:15 - 18:00 21:15 - 22:00
Awards and closing remarks Zoom 10:00 - 10:20 14:00 - 14:20 18:00 - 18:20 22:00 - 22:20
Demo presentation - repeat Zoom 10:25 - 10:40 14:25 - 14:40 18:25 - 18:40 22:25 - 22:40
Poster session C – room K-N 10:30 - 12:00 14:30 - 16:00 18:30 - 20:00 22:30 - 24:00

Block A

Session Location Time      
    US West coast Punta Cana London China
    (UTC-8) (UTC-4) (UTC) (UTC+8) (Nov 21)
Keynote speaker 2 – Roger Levy Zoom 15:00 - 15:45 19:00 - 19:45 23:00 - 23:45 07:00 - 07:45
Poster session A K-N 16:00 - 17:30 20:00 - 21:30 00:00 - 01:30 08:00 - 09:30
Oral session 3 Zoom 17:30 - 19:45 21:30 - 22:45 01:30 - 02:45 09:30 - 10:45
Keynote speaker 3 – Idan Blank Zoom 20:00 - 20:45 23:00 - 23:45 03:00 - 03:45 11:00 - 11:45

Invited speakers

Idan Blank, UCLA

Understanding NLP’s blackbox with the brain’s blackbox and vice versa

This talk will propose a bi-directional link between artificial and biological language processing mechanisms, demonstrating that each can be used as a tool for studying the other. First, I will ask: given what we know about language processing in the human brain and mind, what would success in artificial NLP look like? Specifically, I will focus on dissociations between language and the rest of high-level cognition to significantly narrow the space of “reasonable expectations” we should pose to language models. Next, I will ask: could state-of-the-art NLP systems provide a decent model of the human brain? Here, I will describe promising work demonstrating that some NLP systems can accurately predict brain responses to linguistic stimuli, and offer initial clues into what might drive such brain-machine correspondence.

Roger Levy, MIT

Evaluating and calibrating neural language models for human-like language processing

With new architectures, larger datasets, and greater computational power, neural language models are getting better and better at the tasks they’re trained for and at offering out-of-the-box representations that can be fine-tuned for high performance in new tasks. But are they getting more and more human-like? Here we use linguistic theory and experimental methods inspired by psycholinguistic research to assess zero- and few-shot performance of contemporary neural models on a range of signature human-like language understanding behaviors. While we find impressive successes by models trained on large quantities of text alone, we find clear advantages for models with a symbolic component when training data scale is small. We also obtain success in calibrate models for more human-like processing. Our results highlight the value of insights from psycholinguistics and cognitive science for neural language models of the future.

Anna Rogers, University of Copenhagen

When BERT plays the lottery, all tickets are winning!

The lottery ticket hypothesis was originally developed for randomly initialized models, but might it also apply to pre-trained Transformers? If the “good” subnetworks exist, can they tell us anything about how BERT achieves its performance?

Workshop description

Neural networks have rapidly become a central component in NLP systems in the last few years. The improvement in accuracy and performance brought by the introduction of neural networks has typically come at the cost of our understanding of the system: How do we assess what the representations and computations are that the network learns? The goal of this workshop is to bring together people who are attempting to peek inside the neural network black box, taking inspiration from machine learning, psychology, linguistics, and neuroscience. The topics of the workshop will include, but are not limited to:

BlackboxNLP 2020 is the third BlackboxNLP workshop. The programme and proceedings of the previous editions, which were held at EMNLP 2018 and ACL 2019, can be found here and here.

The call for papers text is available here.

Shared Interpretation Mission

BlackboxNLP 2020 will include a shared interpretation mission. Details available here.

Paper submission

We accept two types of papers

Both papers and abstracts should follow the official EMNLP 2020 style guidelines and should be submitted via softconf:

Accepted submissions will be presented at the workshop: most as posters, some as oral presentations (determined by the program committee).

Dual submissions and preprints

Dual submissions with the main conference are allowed, but authors must declare dual submission by entering the paper’s main conference submission id. The reviews for the submission for the main conference will be automatically forwarded to the workshop and taken into consideration when your paper is evaluated. Authors of dual-submission papers accepted to the main conference should retract them from the workshop by September 20.

Papers posted to preprint servers such as arxiv can be submitted without any restrictions on when they were posted.

Camera-ready information

Authors of accepted archival papers should upload the final version of their paper to the submission system by the camera-ready deadline. Authors may use one extra page to address reviewer comments, for a total of nine pages.


Afra Alishahi

Afra Alishahi ( is an Associate Professor of Cognitive Science and Artificial Intelligence at Tilburg University, the Netherlands. Her main research interest is developing computational models for studying the process of human language acquisition. Recently she has been studying the emergence of linguistic structure in grounded models of language learning. She has chaired CoNLL 2015, and organized the EACL Workshop on Cognitive Aspects of Computational Language Acquisition in 2009, and co-organized the first edition of BlackboxNLP.

Yonatan Belinkov

Yonatan Belinkov ( is an Assistant Professor at the Technion Department of Computer Science. He has previously been a postdoc at the Harvard School of Engineering and Applied Sciences (SEAS) and the MIT Computer Science and Artificial Intelligence Laboratory (CSAIL). His research focuses on interpretability and robustness of neural network models of human language. His research has been published at venues such as ACL, EMNLP, NAACL, TACL, ICLR, and NeurIPS. He serves or has served as an area chair for ACL, EMNLP, and CoNLL, and co-organized the second edition of BlackboxNLP. His PhD dissertation at MIT analyzed internal language representations in deep learning models, with applications in machine translation and speech recognition.

Grzegorz Chrupała

Grzegorz Chrupała ( is an Associate Professor at the Department of Cognitive Science and Artificial Intelligence at Tilburg University. His research focuses on computational models of language learning from multimodal signals such as speech and vision and on the analysis and interpretability of representations emerging in multilayer neural networks. His work has appeared in venues such as Computational Linguistics, ACL, EMNLP and CoNLL. He has served as area chair for ACL, EMNLP and CoNLL and he co-organized the first two editions of BlackboxNLP.

Dieuwke Hupkes

Dieuwke Hupkes ( is a Postdoc at the University of Amsterdam, supported by the ELLIS society. The main focus of her research is understanding how neural networks can understand and learn the structures that occur in natural language. Developing methods to interpret and interact with neural networks has therefore been an important area of focus in her research. She authored 5 articles directly relevant to the workshop, one of them published in a top AI journal (Journal of Artificial Intelligence), and she is co-organizing a workshop on compositionality, neural networks, and the brain, held at the Lorentz Center in the summer of 2019.

Yuval Pinter

Yuval Pinter ( is a PhD student at Georgia Institute of Technology. His main focus is on word-level representations in deep learning systems. He authored two papers on the topic of NLP neural model interpretation in 2019, including one at BlackboxNLP. In addition to regularly serving on program committees for NLP and AI venues, he co-organized the TREC LiveQA competition for its three years of existence (2015–2017), and served as publicity and social media co-chair at NAACL 2019.

Hassan Sajjad

Hassan Sajjadd ( is a research scientist at the Arabic Language Technologies group, Qatar Computing Research Institute - HBKU. His recent research focuses on developing methods to analyze and interpret neural network models both at the representation-level and at the individual neuron-level. His work on the analysis of deep models is recognized at various prestigious research venues such as ACL, NAACL, ICLR, and AAAI.

Program committee

Anti-Harassment Policy

BlackboxNLP 2020 adheres to the ACL Anti-Harassment Policy.