This presentation was given at various events in June 2017 on the current status of Neural Machine Translation development at Iconic.
Rule based, statistical, hybrid, neural - at the end of the day it's all machine translation. At Iconic, we've been "doing neural" for over 12 months in various guises but, frequently, we find that our clients don't care what we use once we get the job done. In these slides, we go through a number of case studies involving MT and show how fit for purpose translations were delivered, combining various different approaches to MT.
4. Impact of Neural MT
Use cases covered by generic MT
Use cases needing custom MTThe
Bar
5. Impact of Neural MT
Use cases covered by generic MT
Use cases needing custom MTThe
Bar
Neural
MT has
raised the
bar
6.
7. Of course.
We’re a team of MT experts. This is a big part of the value that
we bring to the table. We’re not just taking open-source tools off
the shelf. We’re innovating, researching, developing new
processes. Same for Neural MT.
e.g. lexically constrained decoding
Neural MT @ Iconic
“Do you ‘do’ Neural MT?”
8. It’s one of the ways.
MT is not a one-size-fits-all technology. What constitutes the best
approach depends on the language pair, domain, use case, and
various other factors. In some cases, the best approach will be
Neural MT, but not yet all the time.
Neural MT @ Iconic
“Is this the way you do MT now?”
9. When it gives the best output!
When you’re customising MT, there are so many things you can
do – different processors, parameters, ways of combining data
and tuning. We try multiple approaches and allow our systems to
use the best one.
Let’s look at some case studies, but first…
Neural MT @ Iconic
“When do you use it then?”
11. Patents Case Study
Average length: 7 words
Average length: 30 words
3 languages Same data sets Client evaluation
Ranking
1. Unusable
2. Poor
3. Adequate
4. Good
5. Excellent
Criteria
90% Adequate or above
0% Unusable
12. Linguist Review
both pass title criteria
only Iconic MT passed
abstracts
Patents Case Study – Chinese to English
31
3938
19
0
5
10
15
20
25
30
35
40
45
Titles Abstracts
Iconic MT Iconic Neural MT
Outcome
Iconic MT deployed in
production
13. Linguist Review
both passed on titles
only NMT passed abstracts
Patents Case Study – Japanese to English
52
33
0
10
20
30
40
50
60
Titles Abstracts
Iconic MT Iconic Neural MT
Outcome
Iconic MT deployed for titles
Neural MT deployed for
abstracts
4
5
4
4
14. Patents Case Study – Korean to English
40
25
0
5
10
15
20
25
30
35
40
45
Titles Abstracts
Iconic MT Iconic Neural MT
Linguist Review
Iconic MT below criteria
Neural MT significantly better
Outcome
Under review!
2
6
4
3
15. Neural MT raises the bar for general purpose MT
but the bar still needs to be tested.
Customisation Case Study
English to French English to Hindi
BLEU 1-TER BLEU 1-TER
Iconic MT 43.0 (+10.4) 55.2 (+7.7) 46.75 (+12.96) 56.4 (+5.5)
GNMT 32.6 47.5 33.79 50.9
Iconic NMT 39.2 50.5 - -
2 languages 1.5M training segments IT content
16. The Iconic Ensemble Architecture™
Neural MT is another
powerful tool in our
arsenal that helps us
deliver best-in-class
machine translation
output
18. MT @ Iconic – what we don’t do
MT MT MT MT
MT
MTMT
MTMTMTMTMT
MTMTMTMTMTMT
The ability to build your own MT
engines with Moses, Phrasal,
OpenNMT, Nematus, Fairseq.
Provide off-the-shelf general
industry engines. There are
some very adequate solutions for
that!
MT
19. Customised expert-built MT, using
the most appropriate tool for the job,
MT or otherwise.
MT @ Iconic – what we do do!
Develop products and solutions that
incorporate machine translation –
not just access to an API.
20. Engage our expert team on an Neural MT project to see if it works for your content
Neural MT – Early Adopter Program
To date
Custom-developments with
some of our closest partners
Now
Inviting early adopters to
expand the range of casesEarly Adopter Program
Direct people to previous talks I’ve given about the impact and history of neural MT:
SUMMARY
it’s promising, but it’s not a one size fits all taking over approach
still looking case by case in short term
long term, wait and see, it’s exciting
Been speaking recently about Neural MT
The (commercial) narrative around neural MT is moving even faster than the pace of development!
We’ve gone very quickly from explaining what it is, to companies showing initial results, to having fully blown production systems producing the best results ever – and in some cases from companies who never even offered MT before. Wow that’s impressive
Our team at Iconic know a little bit about MT but you can call me a cynic if we take the approach of still trying to manage expectations while we all learn a little more about Neural MT.
Much of it’s still marketing
Still needs to be contextualised
This was always the case with MT, and that doesn’t change with Neural MT
Short-term impact | Long-term prospects
Ultimately just another type of MT so we’re still have a lot of the same issues, and some new issues
Whether that custom MT is neural, or SMT, or hybrid, STILL depends. It’s still to be judged on a case by case basis
Ok, so let’s talk about our approach neural MT at Iconic, but we’ll get some key question out of the way up front.
Questions we're asked frequently as a provider of machine translation
This is what we've been doing for many years now
Fancy way of saying we can apply terminology
With Moses, the way ppl do MT, a lot of this was out of the box. Now we have to implement it ourselves so that will separate the wheat from the chaff in terms of software.
Beam size, phrase length, distortion limit – now training epochs, vocab size, number of hidden layers
Before we talk about the HOW we do it (which is less important as I’ll point out) let’s look at what we’ve done so far with some case studies
REAL RESULTS FROM THE FIELD
A lot of ongoing patent work with some of our biggest clients which is an ideal starting point for us to test our production metal in this area
TEST NEURAL MT ON REAL USE CASES AND CRITERIA
“Baseline” Iconic MT here is
WHAT: Chinese – pre-ordering system, mature 3 years, auto post-editing
OUTCOME: Use Iconic where possible, because it’s quicker to retrain and more control
WHAT: Japanese – syntax based pre-ordering system, transliteration and script normalisation
Interesting, because we were doing A LOT of development before Neural MT but it was hard to make big improvements with constraints on data
OUTCOME: Use Iconic where possible, because it’s quicker to retrain and more control in the short term
Korean project Interesting, because it wasn’t something we had in production before Neural MT – mainly because it was so hard.
WHAT: Korean – hierarchical system
It’s a live one so the results are actually still with the client – LET’S SEE NEXT WEEK, TAUS AND LOCWORLD!
Internal QA on the output suggest the automatic scores are generous to Iconic MT. The Neural output is significantly better
Still ongoing. Iconic Neural MT engines in building, but this is the baseline we establish!
CONCLUSIONS HERE:
Customised Iconic MT is better than general neural MT for 2 very different languages
Again, even customised NMT not as good
So, we will use what we can where’s it’s best.
NEURAL MT CUSTOMISATION NOT YET VERY POWERFUL ACROSS THE BOARD.
WE’RE STARTING TO GET AN INTUITION WHERE IT WILL HELP AND NOT.
Where does NEURAL FIT IN?
HOWEVER, it’s all well and good looking being the curtain and how we do this but at the end of the day, a particular quote springs to mind….
This was a quote from a prospective client on a call last month. I was in the midst of explain what we use in which cases, where neural MT fits in, and they interrupted me as said “I don’t give a damn what you use”. I thought for a second as I was disrupted from my flow and thought, he’s right you know. I’m telling people to leave it to us, and that’s what he wants.
Why does it matter how the translations are produced? Does it really matter if it’s neural or not – once it’s guaranteed as the best of what we could achieve, including using neural. I don’t think it does to anyone!
But because you’ve decided to watch, I’ll give you some insight into what we do and don’t do
- With SMT/NMT, whatever, this it will get you so far before more expertise is required. This is the case now more than ever with NMT and how experimental it is.
- Not trying to provide off the shelf generic engines. Never done that. That’s the realm of Google. You’ll find it’s quite good for the general use case! Good basis for customisation.
We’re doing what we’ve always done!
EDISCOVERY
REGULATORY COMPLIANCE
WHERE MT IS A PART OF A BROADER SOLUTION
To close the loop on this story with Iconic and NEURAL MT
We’re learning, fast, and the more REAL opportunities there are the better for everyone.
Let us know if you have any questions or if it’s something you’d like to explore.
THANKS!