A Silly "Name" Generator

tl;dr: It looks like names aren't well modeled as a Markov process, but you can install my R package that does model names as Markov processes and mess around with it.

I don't know how I wound up writing a "name" generator yesterday, but I did. And now it's an R package (just on github for the moment), so you can play around with it too (https://jofrhwld.github.io/nameGenerator/).

I was messing around with some other research questions when I decided to see what would happen if I tried to model given first names as a Markov process. Here's a picture of a Markov chain that contains just the letters of my own name: Joe.

First, you start out in a start state. Then, you move, with some probability, either to the o,  j, or e character. Then, you move, with some probability, to one of the other states (one of the other letters or end), or stay in the same state. In this figure, I've highlighted the path that my own name actually takes, but there are actually an infinite number of possible paths through these states, including "names" such as "Eej", "Jeoeojo", "Jojoe", etc.

A Markov chain for all possible names would look a lot like this figure, but would have one state for every letter. Now, I keep saying that you move from one state to the next with "some probability," but with what probability? If you have a large collection of names, you can estimate these probabilities from the data. You just calculate for each letter what the probability is of any following letter. So for the letter "j", you count how many times a name went from "j" to any other letter. For boys names in 2013, that looks like this table.

from	to	count
j	a	176452
j	o	84485
j	e	26118
j	u	25616
j	i	1920
j	h	425
j	r	121
j	c	118
j	d	98
j	end	55
...	...	...

As it turns out, a whole bunch of name data is available in the babynames R package put together by Hadley Wickham. So, I wrote a few functions where it estimates the transition probabilities from the data for a given year (from 1880 to 2013) for a given sex, and then generates random "names", or just returns the most probable path through the states. How often does this return a for real name? Sometimes, but not usually. For example, the most probable path through character states for boys born in 1970 is D[anericha] with that"anericha" bit repeating for infinity. For boys born in 1940, it's just an infinite sequence of Llllllll...

So, that introduces a problem where the end state is just not a very likely state to follow any given letter, so when generating random names from the Markov chain, they come out really really long. I introduced an additional process that probabilistically kills the chain as it gets longer based on the probability distribution of name lengths in the data, but that's just one more hack that goes to show that names aren't well modeled as a Markov chain. 

Here's a little sample of random "names" generated by the transition probabilities for girls in 2013:

• Elicia
• Annis
• Ttlila
• Halenava
• Amysso
• Menel
• Seran
• Pyllula
• Paieval
• Anicrl

And heres a random sample of "names" generated by the transition probabilities for girls in 1913:

• Lbeana
• Peved
• Math
• Bysenen
• Viel
• Lelinen
• Jabbesinn
• Mabes
• Drana
• Lystha

The feeling I get looking at these is that they don't seem particularly gendered, even though there are clear gendered name trends. They don't even seem like they're from different times from each other. A lot of them aren't even orthographically valid. I don't know how they'd perform on "name likeness" tasks, but I don't even know what the point of doing such a task would be, since the Markov process has already failed at being a good model of names.

Maybe there's a lesson to be learned from the Markov process' failure to model names well, but for me it wound up being a silly diversion.

---

Update: I've now updated the package to generate names on bigram -> character transition probabilities. The generate_n_names2() function generates things that look more like names. It's kind of fun!

10 generated girl names based on 2010 data:

• Rookenn.
• Lilein.
• Hayla.
• Dailee.
• Bri.
• Samila.
• Abeleyla.
• Eline.
• An.
• Rese.

10 generated boy names based on 2010 data:

• Briah.
• Dason.
• Jul.
• Messan.
• Kiah.
• Jax.
• Se.
• Frayden.
• Dencorber.
• Gel.

America's Ugliest Accent: Something's ugly alright.

I should really blog more often, instead of just when I feel compelled to slap down some nonsense, because the general tone of Val Systems turns towards scolding and away from my genuine positive passion for linguistics. That said, guess what I'm doing in this post!

If you read past the headline, it gets even worse. I won't always reply to examples I find of gross linguistic discrimination like this, because if I did it'd be a full time job. But I noticed that in the introduction they'd linked to a New York Times column that references a paper that I co-authored on the Philadelphia dialect. I didn't think the NYT column was appropriately respectful, and I said so on Language Log at the time.

The NYT columnist wasn't too happy about what I said, but I feel that I have an ethical obligation to the people who invite us into their homes and are generous with their time and stories, to provide them with a vigorous public defense if their communities and the way they speak are ridiculed as a result. Moreover, language shaming pieces like this Gawker tournament only poison the waters for future sociolinguistic research, especially if our names as researchers are attached onto them in some way.

And as I was writing up some notes for this response, and followed more links from the Gawker pieces, I was really shocked by how many articles they've linked to that are popular writeups of sociolinguistic research, usually including interviews with one or more sociolinguists! It's like half my facebook friends list in there! It feels so defeating to see these generally positive articles and interviews utilized to prop up an exercise as ugly and mean spirited as this one.

But what's the harm...?

Anticipating some reactions to this post, no, I'm not some grey humorless lump. But just because something is framed as a game doesn't make it fun, and it doesn't make it funny. For example, take Gawker's paragraph about New Orleans:

New Orleans is a steaming, fetid stew of aural bile, home to everything from the deep Cajun bayou accent to the Yat dialect, which derives from Irish, French, German, and even Italian into one completely incomprehensible mess. You need only watch this clip on the number of ways residents pronounce the city's name and neighborhoods and read this excellent article on the hodgepodge of New Orleans' accents to see how varied, and uniformly ugly, it all is.

lol?

Even if there was some intrinsic humor to that paragraph, instead of just raw nastiness, it would be important to reflect on who we're making fun of, and why. And on that count, I think the fella who compiled the Yat dictionary summarizes it well:

It's a working class language, probably, is what it amounts to.

And of course, that's what linguistic discrimination is really about. Maybe it's not always about class, but it's never really about language. It's about the kind of people who speak it. Predictably, the kinds of accents and languages which get dumped on the most, and get branded the "ugliest," always wind up being spoken by socially disadvantaged people. What exactly did this woman in particular do to deserve having a candid video of her slapped up on Gawker as an example of just how "ugly" the Chicago accent is? She works in a warehouse supermarket, that's what.

And this isn't a consequenceless game either. "America's Ugliest Accent Tournament" just puts a laughing face on a serious problem of discrimination that has economic and personal consequences for real people. To choose one example I'm familiar with, Anita Henderson did a study where she surveyed hiring managers in Philadelphia, playing them tapes of potential job applicants, and asked them to rate them on their job suitability. The topline summary from the abstract:

Those who sound Black are rated as less intelligent and ambitious and less favorably in job level.

In her textbook English with an Accent, Rosina Lippi-Green sums up my own opinion on the matter, but I've added some emphasis.

If as a nation we are agreed that it is not acceptable or good to discriminate on the grounds of skin color or ethnicity, gender or age, then by logical extension it is equally unacceptable to discriminate against language traits which are intimately linked to an individual's sense and expression of self. 

How's this different from these other examples?

A few of the supporting links from the Gawker piece are personal websites that are called "How to talk [City]" or "The [Dialect] Dictionary," put together by enthusiastic speakers of the area themselves. They tend to have a self deprecating tone, so isn't that similar to the Ugliest Accent Tournament? It sure as hell isn't! First of all, even if those personal sites do have a poking fun tone, the fact is the dialect must be important to the person putting together the site, or else they wouldn't have spent the time documenting it! Their self deprecating tone could either be due to the general difficulty of expressing seriously how important a topic is to you, or to their internalized linguistic insecurities driven by things like America's Ugliest Accent Tournament. Moreover, Gawker is a really large media organization, and should be taken to task if only due to their profile and influence.

Sociolinguists ask people what they think about accents and dialects too. It's a subfield sometimes called Perceptual Dialectology. Isn't that kind of the same? Don't even start! A goal of sociolinguists is to understand the social landscape of language as well as we can, and that includes people's sometimes crummy attitudes about it. But if we have a goal, it's to critique those attitudes, not revel in them in some kind of user engagement experiment so that we can go cash in our pageviews with advertisers.

The Fundamental Sociolinguistic Outlook

Real quick, let's contrast the overall tone of the America's Ugliest Accent Tournament with what I would call the fundamental sociolinguistic outlook on speakers. I think Bill Labov summed it up nicely at the end of his 2009 Haskins Prize lecture (go listen to it if you haven't already).

Versus Gawker

No matter who you are, you all sound disgusting.

What to do about it?

I for one will be writing a polite e-mail to Gawker asking them to remove the link that references my research, and to avoid linking to anything that references my research in the future. I'd encourage anyone else whose research they mentioned to do the same.

The new iOS Health app is disappointing

I've been getting into using a few different health tracking apps, and have been getting tired of needing to punch the same data into 3 different places every time I step on a scale. So, I was reasonably excited about the new Health app in iOS8, which would act as one central repository for this information that the individual apps could pull from. The fact that release of the HealthKit API has been delayed, meaning my 3 different health apps can't access the Health data yet, is disappointing, but I'm pretty patient about these things.

However, the Health app itself is really disappointing all on its own. It is not a success of data reporting or visualization. For example, here is what the record of the number of steps I've taken each day for the past month looks like.

So, riddle me this: How many steps did I take yesterday? What was the date that I took the most steps? What day of the week was that weird dip? Not only are answers to basic questions like these not "glanceable," they are totally inaccessible. There is, in fact, no way within the Health app to find these answers, but back to that later.

Let's get a bit more detailed. What is the range of the y-axis. It looks like the bottom horizontal like corresponds to 1,500 steps. That's already a questionable data reporting decision. It should probably correspond to 0 steps. How about the top of the y-axis range? The top horizontal like looks like it corresponds to 13,951 steps, but I'm actually pretty sure that is the maximum number of steps in this data. But the maximum data point doesn't touch the top line?

But let's talk about how Apple really failed to meet baseline expectations with these graphs. When I realized I couldn't read the data precisely off the graph, my first instinct was to drag my finger across the line, assuming that more detailed contextual data would pop up. Sort of like how this Google Ngrams graph works. It should even work on mobile if you tap on it. Or, take this excellent bit of interactive visualization from the New York Times Upshot blog. Or any line graph out there with any bit of polish. Users are more or less trained by this point that hovering over line graphs activates some kind of additional contextual information, whether it's more detailed labeling, brushing, or like that NYT visualization, additional graphs! So you might expect that on the baddest touch screen device ever in the world (as Apple would have us believe), there's going to be some wild and crazy touch interaction, pinch-to-zoom pizzazz. Or at least it might have the same baseline functionality as some silly web widget that I can embed in my blog.

No such luck, and the data viz nerd in me sees this as one of the biggest missed opportunities I've seen in a while. It is just a static image, with some minimal transition animations when you switch between different time scales. If you tap on the graph, you get taken to the raw data, which looks like this.

As far as I can tell, this is the really raw data offered up by the motion co-processor. Ludicrously, you can select and delete any individual bout of steps. So, if I felt that actually, one of the 8 groups of steps logged all in the minute of 10:08 AM was inaccurate, I could delete it!

What really frustrates me about the fact that I can see this data is that I can't touch it. Data at this granularity is pointless other than to show off the fact that there's a lot of it. It needs to a little bit aggregated before it gets interesting. And the fact that I don't like the Health visualizations as it is, I'd really go to town on this raw data. But conspicuously absent here is any export utility. I can look at, but not touch my own data. I guess I also couldn't access the data before iOS8, but they didn't waggle it tantalizingly in front of my nose like this!

So sure, maybe someone will make a third party app that will access the data from the HealthKit API and allow me to export it from there. As if what I'm really dying to do is clutter up my phone with an inevitably junky ad riddled app that contributes functionality that really should've been there in the first place.

To sum up, the static figures are poorly designed and minimally informative, but static figures are hardly what I would expect from a corporate entity like Apple anyway. On top of that, waving this raw data in my face is equal parts useless and infuriating.