The NLP Cypher | 01.03.21

Author(s): Ricky Costa

Originally published on Towards AI.

NATURAL LANGUAGE PROCESSING (NLP) WEEKLY NEWSLETTER

The NLP Cypher U+007C 01.03.21

A New Era

Hey Welcome back, you made it! Now, let us begin 2021 on the right path with an impromptu moment of customer service by Elon Musk:

FYI

If you haven’t read our Mini Year Review, we released it last week while everyone was on holiday U+1F62C. Per usual, if you enjoy the read please give our article a U+1F44FU+1F44F and share it with your friends and enemies!

Now, let’s play a game. Let’s say we have all 7,129 NLP paper abstracts for the entire year of 2020. And now we run BERTopic U+1F447 on top of those abstracts for some topic modeling to find the most frequent topics discussed.

MaartenGr/BERTopic

BERTopic is a topic modeling technique that leverages U+1F917 transformers and c-TF-IDF to create dense clusters allowing…

github.com

What do we get?

1. speech-related
2. bert-related
3. dialogue-related
4. embeddings-related
5. graphs-related

For a more detailed readout of the topics U+1F447

A Pile of 825GBs

The Pile dataset, an 800GB monster of English text for language modeling. U+1F440

The Pile is composed of 22 large and diverse datasets:

paper

The diversity of the dataset is what makes it unique and powerful for holding cross-domain knowledge.

As a result, to score well on the Pile BTB (Bits per Byte) benchmark a model should

…“be able to understand many disparate domains including books, github repositories, webpages, chat logs, and medical, physics, math, computer science, and philosophy papers.”

The dataset is formatted in jsonlines in zstandard compression. You can also view more datasets on The Eye U+1F441 here:

The Pile

The Pile

The Pile is a 825 GiB diverse, open source language modelling data set that consists of 22 smaller, high-quality…

pile.eleuther.ai

The U+1F441

Index of /public/AI/pile_preliminary_components/

The Eye is a website dedicated towards archiving and serving publicly available information. #opendirectory #archive…

the-eye.eu

Domain Shifting Sentiment on Corporate Filings

Corporations are adapting to NLP models that listen in on filings and other financial-related disclosures. According to a new study, corporations are choosing their words wisely in order to fool machines so they are able to reduce the negative sentiment in their statements.

Paper:

How to Talk When a Machine is Listening: Corporate Disclosure in the Age of AI

Founded in 1920, the NBER is a private, non-profit, non-partisan organization dedicated to conducting economic research…

www.nber.org

ML Book Drops U+1F4DA

This week, a couple of ML book prints dropped from well known authors in machine learning. The first is from Jurafsky and Martin’s Speech and Language Processing’s book with new chapters/updates:

Highlights:

-new version of Chapter 8 (bringing together POS and NER in one chapter),

-new version of Chapter 9 (with Transformers)

-Chapter 11 (MT)

neural span parsing and CCG parsing moved into Chapter 13 (Constituency Parsing) and Statistical Constituency Parsing moved to Appendix C

new version of Chapter 23 (QA modernized)

Chapter 26 (ASR + TTS)

Speech and Language Processing

new version of Chapter 8 (bringing together POS and NER in one chapter), new version of Chapter 9 (with Transformers)…

web.stanford.edu

Also Murphy’s Probabilistic Machine Learning draft made the rounds this week. And there’s code along with it! Enjoy.

https://probml.github.io/pml-book/book1.html

code:

probml/pyprobml

Python 3 code for my new book series Probabilistic Machine Learning. This is work in progress, so expect rough edges…

github.com

Open Library Explorer

There’s a new way to explore the Internet Archive for awesome content.

The Open Library Explorer! A new way to browse the Internet Archive

Are you looking for a change of pace this holiday season? How about some reading? Now I'm sure you are all trying to…

datahorde.org

Quantum Ad-List

Someone built U+1F447 as a way to block ads U+1F923.

“Made an AI to track and analyze every websites, a bit like a web crawler, to find and identify ads. It is a list containing over 1,300,000 domains used by ads, trackers, miners, malwares.”

The Quantum Alpha . / The Quantum Ad-List

With over 800000 blocked domains used by ads that my magnificent AI put up together. The AI is like a loyal dog…

gitlab.com

Repo Cypher U+1F468‍U+1F4BB

A collection of recent released repos that caught our U+1F441

LayoutLM V2

Microsoft released the 2nd version of their document understanding language model LayoutLM. If you are interested in SOTA w/r/t document AI tasks. Follow this repo!

microsoft/unilm

December 29th, 2020: LayoutLMv2 is coming with the new SOTA on a wide varierty of document AI tasks, including DocVQA…

github.com

WikiTableT

A large-scale dataset, WikiTableT, that pairs Wikipedia sections with their corresponding tabular data and various metadata.

mingdachen/WikiTableT

Code, data, and pretrained models for the paper "Generating Wikipedia Article Sections from Diverse Data Sources" Code…

github.com

ShortFormer

Shortformer model shows that by *shortening* inputs, performance improves while speed and memory efficiency go up. It uses two new techniques: staged training and position-infused attention/caching.

ofirpress/shortformer

This repository contains the code for the Shortformer model. This file explains how to run our experiments on the…

github.com

ExtendedSumm

An extractive summarization technique that observes the hierarchical structure of long documents by using a multi-task learning approach.

Georgetown-IR-Lab/ExtendedSumm

This repository contains the implementation details and datasets used in On Generating Extended Summaries of Long…

github.com

NeurST

NeurST aims at building and training end-to-end speech translation.

From the TikTok folks at Bytedance:

bytedance/neurst

NeurST aims at easily building and training end-to-end speech translation, which has the careful design for…

github.com

TabularSemanticParsing

Model used in cross-domain tabular semantic parsing (X-TSP). This is the task of predicting the executable structured query language given a natural language question issued to some database.

salesforce/TabularSemanticParsing

This is the official code release of the following paper: Xi Victoria Lin, Richard Socher and Caiming Xiong. Bridging…

github.com

AraBERTv2 / AraGPT2 / AraELECTRA

AraBERT now comes in 4 new variants to replace the old v1 versions.

aub-mind/arabert

This repository now contains code and implementation for: AraBERT v0.1/v1: Original AraBERT v0.2/v2: Base and large…

github.com

Reasoning over Chains of Facts with Transformers

Model retrieves relevant factual evidence in the form of text snippets, given a natural language question and its answer.

rubencart/LIIR-TextGraphs-14

This repository contains the implementation for our submission to the TextGraphs-14 shared task on Multi-Hop Inference…

github.com

Dataset of the Week: DECODE Dataset

What is it?

A conversational dataset containing contradictory dialogues to study how well NLU models can capture consistency in dialogues. It contains 27,184 instances from 4 subsets from Facebook’s ParlAI framework.

Sample

Where is it?

Contradiction

A study on contradiction detection and non-contradiction generation in dialogue modeling. The paper can be found here…

parl.ai

Every Sunday we do a weekly round-up of NLP news and code drops from researchers around the world.

For complete coverage, follow our Twitter: @Quantum_Stat

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI