NLP News Cypher | 01.12.20

Source: Deep Learning on Medium

Before we start, today’s column was inspired by the anthemic 1966 song “ Wild Thing” by The Troggs. Video Below 👇👇

Video starts with the band standing, instruments in hand, on what looks like a hallway of some kind. They then follow a femme fatale through a door into what looks like a back room but turns out they are actually in a middle of a subway station. 🤯🤯

Lead singer does rattle snake head movement the whole time. 🤟🤟

declassified

This Week:

GPT-2 for Tweeting [What you just read]

Neural Module Network Effects

Too Many Recaps, Not Enough Time

Lex’s Vision is 2020

Time for a Fireside Chat?

Reading Comprehension Evaluation Server

Using BERT for NLU

Dataset of the Week: AQuA

Neural Module Network Effects

Nitish Gupta et al. introduces a Neural Module Network model that is able to reason over a paragraph symbolically (arithmetic, sorting, counting) on numbers and dates. It also achieves SOTA on a subset of the DROP dataset.

According to source, code is dropping soon…

Paper:

URL

Too Many Recaps, Not Enough Time

Every big tech company’s AI research arm has come out with a “Year in Review.” This past week it was Facebook and Google’s turn.

My favorite blog post (from Facebook AI’s review) discussed the challenges of open-domain dialogue:

Facebook:

Google:

Me:

Lex’s Vision is 2020

I remember watching Lex’s 2019 video (seen here) and really enjoying it. Well, he has returned. And BTW, NLP gets a big shout-out. Transformers are kind of a big deal. Anyway, lucid recap of the current state of AI across NLP and Computer Vision.

Time for a Fireside Chat?

Wasn’t aware there was a compendium for this. But Microsoft Research shared a collage of various video interviews with the industry’s thought leaders.

Reading Comprehension Evaluation Server

They call it ORB (Open Reading Benchmark). You drop a single question answering model into ORB’s server and it evaluates on several reading comprehension datasets. When submitting your model, they require a docker image that will run on their VM with 4 vCPUs, 1 P100 GPU, and 26GB RAM for eval.

Using BERT for NLU

A fellow named Olivier Grisel fine-tuned BERT to convert an English user query into a representation for handling NLU on task-oriented dialogue. It was fine-tuned on SNIPS, a voice assistant dataset. The project was partly based on the Alibaba paper: https://arxiv.org/pdf/1902.10909.pdf.

Below is an example for intent classification/slots filling on a query:

>>> show_predictions("Book a table for two at Le Ritz for Friday night!",
... tokenizer, joint_model, intent_names, slot_names)
____________________________________________________________________## Intent: BookRestaurant
## Slots:
Book : O
a : O
table : O
for : O
two : B-party_size_number
at : O
Le : B-restaurant_name
R : I-restaurant_name
##itz : I-restaurant_name
for : O
Friday : B-timeRange
night : I-timeRange
! : O

Notebook:

Colab:

Dataset of the Week: AQuA

We’re doing something new, from now on, we’ll highlight an NLP dataset every week.

Ok… back to AQuA… aka Algebra Question Answering with Rationales.

What is it:

“Algebraic word problem dataset, with multiple choice questions annotated with rationales.”

Sample:

"question": "A grocery sells a bag of ice for $1.25, and makes 20% profit. If it sells 500 bags of ice, how much total profit does it make?",
"options": ["A)125", "B)150", "C)225", "D)250", "E)275"],
"rationale": "Profit per bag = 1.25 * 0.20 = 0.25\nTotal profit = 500 * 0.25 = 125\nAnswer is A.",
"correct": "A"

Where is it?