00:00:11.660
So can everyone see the screens? I'm going to try to do a live demo because, you know, I like living on the edge a little dangerously.
00:00:17.910
I'm Aja Hammerly, but I go by Sacre Miser in most places on the Internet. Tigra Miser was taken on Twitter, so it's the Sacre Miser.
00:00:24.000
I love it when people tweet me questions or comments, or even if they think I'm completely wrong about something during talks.
00:00:30.689
My phone is somewhere over there, and I believe Twitter is turned off on the laptop, so you all won't see it anyway. Feel free to go nuts!
00:00:37.890
I blog once a week at SacreMiser.com, covering various topics including DevOps, Ruby, and the art of speaking.
00:00:43.110
I also work at Google on the cloud platform, so if you're interested in Google Cloud, Kubernetes, or the other things we do like machine learning APIs, I'm happy to answer questions.
00:00:48.710
I have a plethora of opinions that I'm more than happy to share with you, so feel free to come find me and we'll chat.
00:00:54.840
And since I work for a large company, I have to say that any code in this talk is copyright Google and licensed under Apache v2.
00:01:05.010
It is currently not on GitHub, but it will all go up there in the next day or two, and I will post about it on my blog and on Twitter when it gets there.
00:01:12.720
Hopefully, since the room is relatively empty, people aren't here by random chance because they didn't like the last talk they went to or walked in a little late.
00:01:18.090
But what is NLP? The acronym stands for Natural Language Processing, which may get you a bit closer to what my talk is about, but it's still somewhat fuzzy for me.
00:01:24.119
Let's go to Wikipedia: Natural language processing is a field of computer science, artificial intelligence, and computational linguistics. It's concerned with the interactions between computers and human natural languages.
00:01:30.180
In particular, it focuses on programming computers to process large data corpora.
00:01:35.250
That's one exceptionally long sentence filled with big words, so here’s the definition I use: teaching computers to understand and ideally respond to human languages.
00:01:41.759
By human languages, I mean things like English, Japanese, American Sign Language, and British Sign Language—basically all the languages humans use.
00:01:48.600
So, to echo millions of middle schoolers everywhere, when are we ever going to use this? And, why should I care?
00:01:55.680
The reason is that bad NLP is already here. Who has ever had to interact with a phone system that says, 'Please say your account number,' and you scream back, 'I didn’t understand you, say it again!'?
00:02:04.170
Who here has logged onto a website and had a chat window pop up with some disembodied figure saying, 'If you need help, I'm here to help you.'?
00:02:10.160
And you wonder if there's really a person behind that or if it's just bad NLP? I had that happen to me last week. It was creepy!
00:02:18.180
And all these things involving natural language processing, frequently bad NLP, are likely involved.
00:02:23.639
However, the promise of NLP is actually better user experiences. I want to live in a world where, instead of teaching people how to interact with computers, we can teach computers how to interact in ways that people already find successful.
00:02:30.180
So, an example of non-ideal NLP that many of you are probably familiar with is this: my favorite slide in the entire talk is 'Computer, tea, Earl Grey, hot.'
00:02:41.759
That was an example of bad NLP; the way I phrased the request was very specific.
00:02:49.179
I wouldn't use that phrase with any other person, but when it was on air, it seemed really futuristic.
00:02:56.120
It turns out we can do better than that. These random things that live in our houses and potentially order food for us are great.
00:03:03.000
There was a talk by Jonah and Julian yesterday, where they were arguing with multiple Alexa devices up here on stage, and I have a Google Home, and I really like it.
00:03:12.750
Depending on the particular brand, they have a relatively large, and in some cases, nearly limitless corpus of voice commands that they can respond to.
00:03:17.640
And, if you really want one of these devices, come to the Google booth; we're raffling one off at 4:30 tomorrow. Stick around until the end of the talk, and I'll show you how to enter!
00:03:25.050
Additionally, I mentioned earlier that we have tech support and phone trees, and these are actually getting a lot better.
00:03:31.140
In my experience, I can actually call my credit card company, say I never received my credit card, and they will send me to the right person who can answer that question.
00:03:37.890
I really hope that this technology improves over the next few years.
00:03:44.200
Now let’s talk about some use cases you may not have thought about; accessibility is one.
00:03:50.940
During the closing keynote yesterday, a point made was that children often use voice interfaces because they can't read yet.
00:03:57.750
I used to work on software for kids, specifically specializing in those with learning disabilities.
00:04:04.220
The ability to use voice interaction or not have to type and spell correctly is fantastic for individuals with dyslexia.
00:04:10.380
It's also great for those with broken arms or individuals who might be holding babies.
00:04:17.850
There are all sorts of use cases where this kind of accessibility matters.
00:04:24.020
In addition to the things that many of you may already think of when I put the word accessibility on the slide, such as for blind people.
00:04:30.054
Blind people can definitely benefit from voice interaction.
00:04:36.230
And the other thing that NLP can help us with is improving our understanding and capacity to analyze large amounts of data.
00:04:42.530
Who here works on an app that has a feedback button somewhere?
00:04:49.350
When I worked in EdTech for kids, the feedback we received via that button was often amazing.
00:04:55.370
Five-year-olds can have the most creative—and sometimes harsh—ways of letting you know they hate your software.
00:05:01.640
And when we had software that became more popular, processing all that feedback was initially really hard.
00:05:07.720
Initially, everyone received every feedback email, and eventually we realized we needed a system to manage it.
00:05:14.090
We wanted to route feedback to the right people, like billing-related inquiries to the billing team.
00:05:20.000
But the complaints about the software needed to go to a 'Things to read on your bad day' folder.
00:05:25.800
And requests for help should go to the team working on that interface.
00:05:32.000
We didn't have a proper way to do this at the time, so we didn't bother.
00:05:36.870
NLP can also assist in various ways. One of my co-workers created a tool called Deep Breath to analyze your emails.
00:05:43.680
If you try to send an email that seems hostile, it tells you, 'You might want to rethink that.'
00:05:50.110
Imagine how many GitHub flame wars could be avoided if everyone had that tool!
00:05:56.520
So, why don't we have that already? Because NLP is hard! Really, really hard.
00:06:02.360
You don't have to believe me; let's talk about why it's hard. First off, English is horrible!
00:06:10.059
Pop quiz: what do you think of when I say the word 'seal'? Raise your hand if you thought of a marine animal.
00:06:16.600
Keep your hand up if you thought of an event where someone signs a contract.
00:06:23.960
Who thought of a musician? Who thought of something completely different?
00:06:30.020
One word can represent many different ideas. This was actually fairly evenly distributed, which is quite interesting.
00:06:35.410
We have our famous 'humma humma' phones, and with so many English teachers around, we have so little time.
00:06:42.810
Additionally, we have words that can be multiple parts of speech. For example, 'love.'
00:06:49.360
Love can be a verb, as in 'She loves her wife,' but it can also be a noun, as in 'Love lasts forever.'
00:06:55.250
English is complicated—don't even get me started on irregular verbs, slang, idioms, and other bits that make a language human.
00:07:02.110
It turns out that English is especially difficult, but all human languages can be quite challenging in similar ways.
00:07:08.180
Even the languages we've tried to manufacture to be less confusing are still complex because humans create idioms and slang.
00:07:15.120
Language evolves and becomes particularly troublesome for computers because there's no formal closed grammar for human languages.
00:07:22.000
While some languages are more irregular than others, grammar for computer languages typically fits on a page or two of printed text.
00:07:29.180
They're generally very regular, with a limited vocabulary; the rules are straightforward.
00:07:35.580
Human languages are messy. For instance, we can say, 'I'm starving,' but that has a remote chance of being true.
00:07:41.700
Or if I say to someone, 'You look freezing,' the probability that’s true is quite low.
00:07:48.240
We exaggerate a lot, and I recently read a cool article stating that the word 'unique' is becoming less unique over the past thirty years.
00:07:54.840
Thirty years ago, the word 'unique' was relatively rare in newspaper articles. Frequently, editors would use 'unusual' instead.
00:08:01.340
But now, it's almost as likely to find 'unique' as you are to find 'unusual,' which is problematic.
00:08:07.840
Another example I like is the word 'literally.' As of two years ago, it doesn't even mean literally anymore!
00:08:14.490
We also have the issues with sarcasm: it's not just humans who struggle, but computers do too.
00:08:20.210
This is an example I used in a previous talk: if I say, 'Sure, I’d love to help you out with your problem,' it can sound sincere.
00:08:28.300
However, if I say, 'Sure, I’d love to do the dishes,' depending on how I say it, the meaning changes.
00:08:34.440
The context and surrounding words really affect how we convey meaning.
00:08:41.820
Despite learning from Douglas Adams's 'The Hitchhiker’s Guide to the Galaxy,' computers are really bad at distinguishing sarcasm.
00:08:47.280
And they struggle even more with generating it. So, I ask again, why is this so hard?
00:08:53.950
I hope I’ve confused you and convinced you, but the main reason is that natural language processing is difficult.
00:09:01.100
Human languages are complicated and always changing. But because we created them, we can simplify everything.
00:09:07.670
So why is NLP hard? Because understanding natural language is hard—humans are tricky!
00:09:14.030
At its core, language is hard for computers due to the intricacies of human communication.
00:09:21.100
However, not all is lost! We've been working on this problem for a really long time.
00:09:27.920
Two famous mathematicians proposed ways to accomplish algorithmic translation between languages.
00:09:36.880
I am of the generation that saw Google Translate become a thing as I was growing up.
00:09:42.630
Automated translation has improved greatly over the past 15 to 20 years, but we have been working on this problem for hundreds of years.
00:09:50.200
Another great example is Alan Turing, who proposed the Turing Test back in 1950.
00:09:57.270
Who here has heard of the Turing Test? A few people. For those who haven't, let’s define it.
00:10:03.130
The Turing Test is a way to measure machine intelligence using natural language.
00:10:09.750
An impartial judge observes a conversation between a human and a machine without knowing which is which.
00:10:15.240
Usually, the conversation occurs via text. If the judge cannot identify which participant is the machine, the machine passes the Turing Test.
00:10:22.010
People have tried to solve this challenge. One of my favorite examples is Eliza, which came out shortly after the Turing Test.
00:10:29.990
If you use Emacs, you can play with Meta-X Eliza in real-time. Since I have the time, let's do that now!
00:10:36.000
So, here we are! Hi, I’m feeling sad! Eliza replies, 'Hi, how do you do? What brings you to me?'
00:10:42.540
I respond, 'I already told you I'm feeling sad.' And Eliza says, 'Why do you say you already told me that you're feeling sad?'
00:10:48.230
And I might say, 'Because I already told you, you stupid machine!' And Eliza would respond, 'Is it because you already told me?'
00:10:58.050
This is a classic example of NLP, and many implementations exist.
00:11:03.360
My project for the plane ride home is to port a Python implementation to Ruby, because, hey, why not?
00:11:10.260
So, typically, chatbots have been around since the late 50s and early 60s, which is somewhat surprising because many think they were invented recently.
00:11:16.900
Now I've rambled on for about 15 minutes without showing you any code, and that's not fair.
00:11:23.650
Code! I like to give a super impractical example in talks so that, hopefully, no one will actually use my examples, but they'll apply the ideas.
00:11:30.440
In RubyConf 2015, I demonstrated sentiment analysis of tweets using emoji scores.
00:11:37.120
Sentiment analysis is a subfield of natural language processing focused on determining if a body of text is generally positive or generally negative.
00:11:44.710
Using emoji, I assigned scores to indicate how positive or negative they were.
00:11:50.860
This was my general scale; I may have partly done this to put the poop emoji on the screen many times!
00:11:57.110
I utilized emoji at the time, and you can watch the video on Confreaks because NLP is super hard.
00:12:05.320
In November 2015, I didn't have the skills to train a model capable of effective natural language processing on the crazy tweets during the conference.
00:12:12.690
However, most people who do machine learning don’t build their own models.
00:12:14.470
Building models requires a lot of time and knowledge since there are many pitfalls.
00:12:21.500
Over the last year, many pre-trained models have been released that you can access via their API.
00:12:28.140
I will use the Google Cloud natural language API for my demo today, rather than my hacky emoji system.
00:12:34.830
We have a gem, which is available on GitHub. Just install 'google-cloud-language,' and it’s currently in alpha.
00:12:42.340
Here's the code I need to perform the sentiment scoring. I'll create a new object and use the 'analyze' method.
00:12:49.060
The 'analyze' method receives a document created from the tweet text. When I call the sentiment method, it returns an object with a score and a magnitude.
00:12:57.290
The score indicates if the sentiment is positive or negative, ranging from -1 to 1. Magnitude indicates how strong that sentiment is.
00:13:04.130
For example, if I get a score of 0.1 and a magnitude of 0.1, that indicates a very neutral tweet.
00:13:10.680
On the other hand, if I have a score of negative 0.9 and a magnitude of 0.1, it's negative but not excessively so.
00:13:18.630
I have a guide about how to interpret those if you're curious because understanding those two things can be difficult.
00:13:25.210
Don’t multiply the score by the magnitude; it doesn't work that way—I’ve already tried!
00:13:32.420
I'm going to skip over how I set up everything. It's a kind of funky nineties cluster pulling data from Twitter.
00:13:39.580
I'd be happy to explain more afterward, so come to my booth where everything is running.
00:13:46.560
I explained it all in detail in my original talk about setting up distributed systems for crazy things.
00:13:54.000
If we have enough time, I’ll do a demo at the end of the talk, but I need your help!
00:14:00.610
I’ll be pulling from that hashtag, so if everyone could take a photo or tweet something, please include a sentiment in your tweet.
00:14:06.740
My tweet has been running pretty good through the conference so far.
00:14:12.930
It’s not surprising—now I’m going to take a drink of water while you all tweet, and I’ll be back in a second!
00:14:18.190
Okay, now that you're all done tweeting, let's talk about grammar for a bit.
00:14:26.800
When this tool came out, I was thinking about how many of us learned how to diagram sentences in grade school.
00:14:33.520
Did you hate that process? I certainly did! It drove me insane to repeat the same thing over and over.
00:14:39.700
If you've never seen this before, this is one form of sentence diagramming.
00:14:46.560
In this diagram, you have a subject connected by a vertical line to a verb, which is also connected by another vertical line to a direct object.
00:14:52.900
All the modifiers go on lines below. Anyone here had to draw diagrams like this? Show of hands!
00:14:58.120
Okay, I'm clearly not alone in my misery.
00:15:03.800
With both methods of diagramming, the verb is at the center, surrounded by everything that modifies it.
00:15:09.640
When I showed this to some of my friends in Seattle at Ruby, they groaned. They hate grammar!
00:15:15.340
Now, I'm returning from that brief side quest about grammar—let’s dive a little deeper.
00:15:21.280
One way we understand words is by labeling them based on their function; this is called parts of speech.
00:15:27.280
Verbs denote actions; you can’t have a sentence without a verb. One type of action could be a verb like 'jump'.
00:15:33.200
However, there are also state-of-being verbs like 'thinking.'
00:15:39.150
Next, we have nouns. A noun is a person, place, or thing—that's strictly true.
00:15:45.040
A person could be Matt or Alan Turing; a place could be the bathroom or Phoenix; and a thing might be a cactus or mountain goat.
00:15:51.170
You can even have nouns that signify abstract ideas, like 'democracy' or 'love.'
00:15:56.750
Then there are adjectives, which describe or modify nouns. They can specify attributes like color, size, or quantity.
00:16:02.720
Additionally, there are determiners, modern grammar calls them. Determiners clarify which noun is being referenced.
00:16:10.440
They include articles like 'a' and 'the,' as well as this and that.
00:16:17.200
Not every determiner is an article, but all articles are determiners—it’s like squares and rectangles.
00:16:22.676
Now, let's talk about sentence structure. You need a verb for a complete sentence.
00:16:29.650
Thus, in the sentence 'The cat eats fish,' the subject is 'cat,' the verb is 'eats,' and the direct object is 'fish.'
00:16:36.080
Side quest complete! Now that we've reviewed the basics of grammar, let's return to our NLP and sentence diagramming.
00:16:42.200
In diagramming a sentence, we begin with the subject, the verb, and the direct object, along with any modifiers.
00:16:49.410
To perform the analysis, I need to identify the part each word plays in the sentence.
00:16:57.000
There are tools for that! Instead of using the sentiment method, I will use the syntax method.
00:17:03.510
This will return a list of tokens along with a wealth of information about each token.
00:17:09.480
For instance, for the word 'cat' in 'The cat eats fish,' I would receive details such as its function in the sentence.
00:17:16.620
The API returns the part of speech, indicating that 'cat' is a singular noun.
00:17:24.340
Interestingly, English lacks many modifiers and can trip up algorithms.
00:17:30.920
If you’re trying to understand parts of speech, it can be enlightening to use it on other languages, like German or Spanish.
00:17:41.100
This token system also includes a 'head token index,' denoting the parent word in relation to the current token.
00:17:47.160
Where 'the' is linked to 'cat', we can see how they are structurally associated.
00:17:53.390
With this information, I can create an ASCII art diagram for a sentence.
00:18:00.600
I can identify the subject of the sentence based on the token label.
00:18:05.970
Next, I find the direct object and string together a more refined ASCII art diagram.
00:18:12.110
However, I was missing one word, so I looked again to identify how the direct object fits into the sentence.
00:18:19.050
The NLP API provides token indices for relationships among tokens.
00:18:25.480
For example, 'the' refers to 'cat' as its main subject, allowing for straightforward diagramming.
00:18:31.470
I simply iterate through all my tokens and print those that refer to the subject.
00:18:38.079
I can produce a basic diagram here, which I'm pretty happy with!
00:18:44.320
Now, let’s step it up with a more complicated sentence to form a more robust visual.
00:18:51.710
Let’s try 'The cat eats fish with a side of milk.' In the analyst’s eye, that adds layers of complexity.
00:18:58.090
Due to the prepositional phrase at the beginning, I can examine how that fits within the structure.
00:19:05.240
The 'with milk' phrase influences how I interpret the entire sentence and shows fluidity in grammar.
00:19:11.950
There are all sorts of practical uses of NLP.
00:19:18.330
We've talked about customer feedback analysis, summarizing content, and creating more accessible products.
00:19:25.200
At this point, I hope I’ve done my job convincing you that there’s great potential here.
00:19:31.750
If you want to get started, the Google Natural Language API is an excellent resource.
00:19:37.790
The first 5,000 requests to each endpoint in text sentiment are free.
00:19:44.280
Once you exceed that, it is reasonably priced per thousand.
00:19:51.450
I also had the Jabberwocky analyzed, and to my surprise, it got the syntax analysis correct.
00:19:59.630
It’s a wonderful opportunity to experiment with technology and see what you're capable of.
00:20:06.820
Dig in and play around! I work for Google, and we have a booth at RailsConf this year.
00:20:14.320
There are live coding labs, cool stickers, and answers at our booth.
00:20:20.480
Also, come grab a dinosaur sticker if you hang out with me afterward!
00:20:26.570
We are also giving away a Google Home—check the link or the whiteboard at our booth.
00:20:33.450
Thank you for your attention! Now I'd be happy to take any questions.
00:20:39.300
Yes, over there; I see a hand raised.
00:20:45.600
The first question was about how the sentence diagramming approach deals with incorrect grammar.
00:20:52.960
I have tried it out with sentences that are missing punctuation and it struggles.
00:20:58.590
Yes, it’s a machine, so machines make mistakes, and they're improving over time.
00:21:05.390
My general rule is that I have bad grammar, i.e., I need a copy editor for my blog.
00:21:12.180
However, it performs well with the common mistakes individuals tend to make.
00:21:19.590
Others may handle less common mistakes better, so there's always room for improvement.
00:21:27.260
The next question was about how the model interprets emojis.
00:21:31.200
That's something I've been testing, and I can show you the current sentiment!
00:21:36.570
The current sentiment is positive, which indicates a majority of positive sentiment in your tweets.
00:21:43.440
Even if some try to skew it negatively, positive sentiments are outweighing them.
00:21:50.890
Lastly, cleaning up data and ensuring high-quality data definitely increases the accuracy.
00:21:58.390
I’ve been pretty lazy and have skipped significant data cleaning so far.
00:22:06.540
But I've been satisfied with results without pre-processing, though it’s something to consider.
00:22:12.600
Another question focused on how the analysis identifies proper nouns, companies, and entities in context.
00:22:20.400
One use case is reviewing SEC filings, identifying key players and significant entities.
00:22:27.720
The sentiment also works well for statements related to personal preferences.
00:22:34.540
Thank you all for a fantastic audience today! Your questions were insightful.
00:22:41.140
You can come get dinosaurs from me anytime. I appreciate your time. Thank you!
