Yonatan Belinkov

#### Issue description The tutorials from ACL 2020 do not have their slides+videos, which were recorded by SlidesLive. #### Steps to reproduce the issue Here's an example: Tutorial 1: https://www.aclweb.org/anthology/2020.acl-tutorials.1/...

This is a list of potential papers and other resources regarding visualization that may be added. - [BertVis](https://github.com/jessevig/bertviz) and [Deconstructing BERT: Distilling 6 Patterns from 100 Million Parameters](https://towardsdatascience.com/deconstructing-bert-distilling-6-patterns-from-100-million-parameters-b49113672f77) and [Deconstructing...

This is a list of potential papers to add that describe challenge sets (Section 4 of the paper, Table SM2 in the supplementary materials/website). - [Grammatical Analysis of Pretrained Sentence...

- [DISCRETE ADVERSARIAL ATTACKS AND SUBMODULAR OPTIMIZATION WITH APPLICATIONS TO TEXT CLASSIFICATION](https://www.sysml.cc/doc/2019/79.pdf) - [Adversarial attacks against Fact Extraction and VERification](https://arxiv.org/abs/1903.05543) - [On Evaluation of Adversarial Perturbations for Sequence-to-Sequence Models](https://arxiv.org/abs/1903.06620) -...

Methods that explain specific predictions: - [On Attribution of Recurrent Neural Network Predictions via Additive Decomposition](https://arxiv.org/abs/1903.11245) - [LS-Tree: Model Interpretation When the Data Are Linguistic](https://arxiv.org/abs/1902.04187) - [Generating Token-Level Explanations for...

- [A Structural Probe for Finding Syntax in Word Representations](https://nlp.stanford.edu/~johnhew/structural-probe.html) - [Introducing Orthogonal Constraint in Structural Probes](https://arxiv.org/abs/2012.15228) - also related to individual dimensions - [Evaluating the Representational Hub of Language...

The Marmot documentation mentions feature-templates for training: ``` Comma separated list, activates individual templates. Default value: "form,rare,affix,context,sig,bigrams" ``` Is there any documentation for the meaning of these templates, what are...

This is information from Bertrand Higy about a problem with DeepSpeech2 Torch version, which is inherited by our implementation: > DeepSpeech2 paper and pytorch implementation use 41x11/21x11 convolutional layers (FxT)...