The NLP Cypher | 05.23.21

Last Updated on July 24, 2023 by Editorial Team

Last Updated on May 24, 2021 by Editorial Team

Author(s): Quantum Stat

NATURAL LANGUAGE PROCESSING (NLP) WEEKLY NEWSLETTER

Overtime

Hey Welcome back, another week goes by and so much code/research has been released into the wild.

Oh and btw, The NLP Index is on ??? , and I want to thank all contributors!

Here’s a quick glimpse at the awesome contributions: A collection of Spanish Medical NLP datasets brought to you by Salvador Lima in Barcelona. ?? Will update the NLP Index with these and other assets by tomorrow.

Cantemist (oncology clinical cases for cancer text mining): https://zenodo.org/record/3978041

PharmaCoNER (Pharmacological Substances, Compounds and proteins in Spanish clinical case reports): https://zenodo.org/record/4270158

CodiEsp (Abstracts from Lilacs and Ibecs with ICD10 codes): https://zenodo.org/record/3606662

MEDDOCAN (Medical Document Anonymization): https://zenodo.org/record/4279323

MESINESP2 (Medical Semantic Indexing): https://zenodo.org/record/4722925

Wav2vec-U: Unsupervised Speech Recognition ?

This new FAIR model doesn’t need transcriptions to learn speech. It just needs unsupervised speech recordings and text. They used a GAN to help discriminate phonemes (sounds of language). While Wav2vec-U doesn’t achieve SOTA on the Librispeech benchmark, it still gets a pretty good score given the fact it didn’t require 960 hours of transcribed speech data. ?

Blog:

wav2vec Unsupervised: Speech recognition without supervision

Code:

pytorch/fairseq

Polars Dataframes ?

If you use dataframes often, you should check out Polars. It’s an awesome dataframe library written in Rust (includes Python bindings). Comes with Arrow support and all of its glory including parquet file and AWS S3 IO support.

pola-rs/polars

Docs:

Polars – User Guide

Universiteit van Amsterdam | Notebooks and Tutorials

The University of Amsterdam has a sweet collection of colab notebooks mixing various domains including GNNs, Transformers and computer vision.

Here’s their TOC:

Tutorial 2: Introduction to PyTorch

Tutorial 3: Activation functions

Tutorial 4: Optimization and Initialization

Tutorial 5: Inception, ResNet and DenseNet

Tutorial 6: Transformers and Multi-Head Attention

Tutorial 7: Graph Neural Networks

Tutorial 8: Deep Energy Models

Tutorial 9: Autoencoders

Tutorial 10: Adversarial Attacks

Tutorial 11: Normalizing Flows

Tutorial 12: Autoregressive Image Modeling

Welcome to the UvA Deep Learning Tutorials! – UvA DL Notebooks v1.0 documentation

KELM | Converting WikiData to Natural Language

Google introduces the KELM dataset in a huge win for the factoid nerds. The dataset is a Wikidata knowledge graph converted into natural language with the idea of using the corpus for improving the factual knowledge in pretrained models! A T5 was used for this conversion. The corpus consists of ~18M sentences spanning ~45M triples and ~1500 relations.

KELM: Integrating Knowledge Graphs with Language Model Pre-training Corpora

Talkin’ about knowledge graphs…

An Introduction to Knowledge Graphs

No Trash Search!

No Trash Search

LabML.AI Annotated PyTorch Papers

Learn from academic papers annotated with their corresponding code. Pretty cool if you want to decipher research.

labml.ai Annotated PyTorch Paper Implementations

Completely Normal (aka not suspect) Task

applicaai/kleister-charity

Repo Cypher ?‍?

A collection of recently released repos that caught our ?

Measuring Coding Challenge Competence With APPS

A benchmark for code generation.

Check out the GPT-Neo results when compared to GPT-2/3, very interesting.

paper

hendrycks/apps

Connected Papers ?

wikipiifed — Automated Dataset Creation and Federated Learning

A repo for automating dataset creation from wikipedia biography pages and utilizing the dataset for federated learning of BERT based named entity recognizer.

ratmcu/wikipiifed

Connected Papers ?

OpenMEVA Benchmark

OpenMEVA is a benchmark for evaluating open-ended story generation.

thu-coai/OpenMEVA

Connected Papers ?

KLUE: Korean Language Understanding Evaluation

KLUE benchmark is composed of 8 tasks:

• Topic Classification (TC)
• Sentence Textual Similarity (STS)
• Natural Language Inference (NLI)
• Named Entity Recognition (NER)
• Relation Extraction (RE)
• (Part-Of-Speech) + Dependency Parsing (DP)
• Machine Reading Comprehension (MRC)
• Dialogue State Tracking (DST)

KLUE-benchmark/KLUE

Connected Papers ?

Contextual Machine Translation

Context-aware models for document-level machine translation. Also includes SCAT, an English-French dataset comprising supporting context words for 14K translations that professional translators found useful for pronoun disambiguation.

Most MT models are on the sentence level, so this is an interesting repo for those looking to go onto the document level.

neulab/contextual-mt

Connected Papers ?

Dataset of the Week: Few-NERD

What is it?

Few-NERD is a large-scale, fine-grained manually annotated named entity recognition dataset, which contains 8 coarse-grained types, 66 fine-grained types, 188,200 sentences, 491,711 entities and 4,601,223 tokens. Three benchmark tasks are built, one is supervised: Few-NERD (SUP) and the other two are few-shot: Few-NERD (INTRA) and Few-NERD (INTER).

Sample (in typical NER format)

Between O
1789 O
and O
1793 O
he O
sat O
on O
a O
committee O
reviewing O
the O
administrative MISC-law
constitution MISC-law
of MISC-law
Galicia MISC-law
to O
little O
effect O
. O

Where is it?

thunlp/Few-NERD

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

---

The NLP Cypher | 05.23.21 was originally published in Towards AI on Medium, where people are continuing the conversation by highlighting and responding to this story.

Published via Towards AI