Transformer: A Novel Neural Network Architecture for Language Understanding | Hacker News Reader

Transformer: A Novel Neural Network Architecture for Language Understanding

research.googleblog.com

280 pointsandrew372619 comments

Very interesting approach, and intuitively it makes sense to treat language less as a sequence of words over time and more as a collection of words/tokens with meaning in their relative ordering.

Now I'm wondering what would happen if a model like this were applied to different kinds of text generation like chat bots. Maybe we could build actually useful bots if they can have attention on the entire conversation so far and additional meta data. Think customer service bots with access to customer data that can learn to interpret questions, associate it with their account information through the attention model and generate useful responses.

No doubt a holy grail for chat bots, but I'll believe it when I see it.

"Even without evidence, everyone should believe it will solve the problem, but I won't believe it will solve the problem until there's evidence."

Is that what you just said? :)

The harder problem, as usual, is to get enough high quality training data for a particular problem domain though.

Maybe as a POC we can try building a bot that generates relevant HN comments given the post and parent comments. Maybe I'm such a bot, how could I possibly know?

A "collection of words in a relative ordering" is "a sequence of words".

They mentioned tensor2tensor, how is this related to another repo: seq2seg:https://github.com/google/seq2seq? Which one is more general?

DeepL (was on HN earlier this week) also uses an attention-based mechanism like this (or at least, with the same intention and effect). They didn't really talk about it but the founder mentioned it to me. The two seem to have independently pursued the technique, perhaps from some shared ancestor like a paper they both were inspired by.

Attention is used in machine translation since 2014 as far as i know (Dzmitry Bahdanau, Kyunghyun Cho, Yoshua Bengio: Neural Machine Translation by Jointly Learning to Align and Translate) in various forms.

The paper linked in the blog was actually published back in June. And as the sibling comment says, people have been trying to use attention mechanisms to augment recurrent networks for some time. The novel idea here seems to be sorting the recurrent network, and just keeping the attention mechanism.

Discussion about DeepL: https://news.ycombinator.com/item?id=15122764

Attention is not new. Everyone uses it (for translation and many related tasks). It's very much the standard right now.

Avoiding recurrent connections inside the encoder or decoder is also not completely new. That came up when people tried to only use convolutions.

Googles Transformer was made public in June 2017, in the paper Attention is all you need, https://arxiv.org/abs/1706.03762, including TensorFlow code, https://github.com/tensorflow/tensor2tensor . Note that the new thing here is that they neither use recurrence nor convolution but rely entirely on self-attention instead, with simple fully-connected layers, in both the encoder and the decoder.

DeepL directly compares their model to Transformer, in terms of performance (BLEU score), here: https://www.deepl.com/press.html

The key to this paper is the "Multi-Head Attention" which looks a lot like a Convolutional layer to me.

Has anyone come across good ML to do arabic-english? Seems to be a complete lack of decent training data.

There's the GALE data provided by LDC (https://catalog.ldc.upenn.edu/):

- GALE Phase 1 Arabic Broadcast News Parallel Text - Part 1 (LDC2007T24) - GALE Phase 1 Arabic Broadcast News Parallel Text - Part 2 (LDC2008T09) - GALE Phase 1 Arabic Blog Parallel Text (LDC2008T02) - GALE Phase 1 Arabic Newsgroup Parallel Text - Part 1 (LDC2009T03) - GALE Phase 1 Arabic Newsgroup Parallel Text - Part 2 (LDC2009T09) - GALE Phase 2 Arabic Broadcast Conversation Parallel Text Part 1 (LDC2012T06) - GALE Phase 2 Arabic Broadcast Conversation Parallel Text Part 2 (LDC2012T14) - GALE Phase 2 Arabic Newswire Parallel Text (LDC2012T17) - GALE Phase 2 Arabic Broadcast News Parallel Text (LDC2012T18) - GALE Phase 2 Arabic Web Parallel Text (LDC2013T01)

Isn't Arabic an official UN language?

I'm a novice when it comes to neural network models, but would I be correct in interpreting this as a convolutional network architecture with multiple stacked encoders and decoders?

Nah, the paper explicitly states that their system is not recurrent nor convolutional:

> To the best of our knowledge, however, the Transformer is the first transduction model relying entirely on self-attention to compute representations of its input and output without using RNNs or convolution.

would something like this work well on mixed/pidgin languages - e.g. Hinglish , which is a mixture of Hindi and English and used in daily vernacular ?

does it mean we don't need gradient descent after all to achieve the same result?

Nope, Transformer is still trained with gradient descent.

Reply on news.ycombinator.com