We’ve spent a lot of time working with natural language processing (NLP) processes and timing lined up for us to try out some fairly straightforward demos of audio transcription and named entity identification. We originally set out to build a model to classify and identify entities in AWS’ release notes – and we’re pretty happy with those results. We’ll post more on this shortly!
From there, we kind of set out to automate a ‘live blog’ of sorts using natural language processing. The idea was that by identifying when the subject of conversation changed by watching the trends in the presentation, we could emulate the ‘something new to say’ trigger we’ve seen with live blogging at keynote presentations. What ensued was a bigger lesson in the amount of effort it takes to take audio data and make it useful text.
In our testing, we found that the newly released Amazon Transcribe streaming solution, even with the custom vocabulary setup, never quite got us to ‘good enough.’ To be fair, there’s a lot of jargon involved but you can check out our vocabulary doc and the AWS documentation on it as well. Additionally, we found some unique problems in processing the transcription output even when we ‘got it right’ with getting acronyms and jargon to be recognized. From odd spacing to word splitting and case sensitivity issues with the NLP process, we
Our NLP Setup…
Just so you can follow along, our setup was really straightforward:
- A custom named entity recognizer (NER) built on top of spaCy’s toolkit.
- The Java sample extended to post the finalized text from Amazon Transcribe to our little NLP entity processing service.
- Some extra processing using TextBlob to help supplement the output of our spaCy NER and to do some quick sentiment analysis.
Run it on a laptop, pipe the output of the ‘livestream’ (YouTube videos of prior re:Invents) into the process and see what happens!
A Good Example of ‘Almost’…
This comes before any feature. And any of us will forever be a number one investment there. Protecting your customers. Your business should be your number one priority. Fortunately, the architect executed to give you a whole set of best practices to to follow. Implement a Stone identity foundation, for example, go for a minimum, perfect for these people. So I know that once we start building our system in general, everybody against room privileges to do everything that theyWerner Vogels – AWS re:Invent 2017 Day 2 Keynote 45:59 to 46:34
See the new H one instances lunch last night, which is perfect for those Get the hardware. You still want the elasticity of the reliability and steal ability of eight of you asked. We now have their metal instances that we now slash Immunity for you. We have the most powerful GPU is out there and three instances. You want a PGA instance? We have that. Not a ray of instances and then what we do is we make all of our excess capacity. Anyone point available to you a spot market. View of applications that can afford to be in a rush. The intermittent where you use the capacity ones available and you don’t win. It’s not spot allows you to save about ninety percent on the price of on demand instances, which is really you…Andy Jassy – AWS re:Invent 2017 Day 1 Keynote 24:10 to 25:00
Almost, but not quite
If you watch the videos (we kept the examples to a minute each) you’ll see something really interesting: the text is really quite close, but there are some big misses. That only compounds when you try to take the output of a process like this and attempt to apply an NER to identify what’s there. Just to put a point on it:
- Andy Jassy was talking about FPGA’s, not golfing
- Werner Vogels might be interested to know I live fairly close to Stone Brewing, but I’m not sure ‘Stone identity’ is a thing
In all, I’m super excited to hear what Andy has to say on Wednesday morning and Werner’s presentations are always awesome. Check out their keynotes via the live stream or on YouTube when they’re released if you’re not able to make it in person.
We aren’t worried
We think this is some pretty awesome stuff, but it’s probably not ready for prime time. In our testing we also saw some ‘less than savory’ language fly through, we thought we’d possibly avoid an embarrassing post or entry that doesn’t quite fit, so while we’ll be doing our tweeting and posting ‘the traditional way,’ look out for natural language processing and AI-derived data analysis from re:Invent and beyond!