The NLP Cypher | 09.05.21
Hey Welcome Back! A flood of EMNLP 2021 papers came in this week so today’s newsletter should be loads of fun! 😋
But first, a meme search engine:
An article on The Gradient had an interesting take on NLU. It describes how a NNs’ capacity for NLU inference is inherently bounded to the background knowledge it knows (which is usually highly limited relative to a human). Although I would add a bit more nuance to this by sharing that this is only a problem for a model that is not localized for its user, meaning a model that wasn’t fine-tuned/prompted (localized) for a specific user. For information that is general and with ground truth i.e. (rain is wet or rain falls down to the ground), the MTP isn’t a big issue with large enough data/model.
I think a bigger issue in NLU (using text only) is when data doesn’t match the complexity of real-world. Meaning there isn’t enough information in the text only modality. Humans by default use a multi-modal approach (text, audio, visual etc.) when interpreting the world around us which helps us with inference. Multi-modal learning can be a viable approach to the MTP problem examples discussed in the article.
For those into document (PDF) parsing 👇. Includes the 2nd version of LayoutLM and also its multi-lingual cousin LayoutXLM.
…And there’s already a repo built on top of these models! 👌
Had previously mentioned the highlights/shorter version on a previous newsletter, now you can get the full dataset:
A long and awesome introduction to graph neural networks.
Holy Moly 🤯
The Compendium contains over 500-topics in ML, and has been written for over 4 years. It’s now offered in an interactive web-based format.
A collection of recently released repos that