2019-12-01

Constituency and Coordination

For Syntax FANBOYS

As any seasoned linguist will tell you, grammar based on folk theory is, well, folksy. And, since most folks are superficial in their linguistic analysis, so, too, are their contrived grammars. At the opposite pole, we have crushing formality. And, since most formalities are constrained, so, too, are their contrived grammars. What lies in between is how I perceive linguistic syntax. The top-down system derives from bottom-up observations. Message parsing has focused on giving a bottom-up method of transformations that preserves semantic value (for the most part). What follows here is formal insights (expressed in constituency grammar) that message parsing invites. And, following that, I aim to tackle this question: How the hell does coordination jive with constituency?

For those who don't read my work (that is, pretty much everybody), here's an example of a message parse with some respective notes:
  1. When did he and why did he do that? : START
  2. When did he do that and why did he do that? : (1) ADD
  3. When did he do that? Why did he do that? : (3) SPLIT
  4. Did he do that at some time? Did he do that for some reason? : (3) MOVE * 2
  5. He did do that at some time. He did do that for some reason. : (4) MOVE * 2
  6. He did do that. He did do that. : (5) CUT * 2
The ADD rule replaces elliptical elements.

The SPLIT rule divides sentences connected by coordinators.

The MOVE rule shifts and replaces constituents to do any of the following:
  • Revert a sentence to a truth-apt form, or 
  • Move a sentence's constituent to a canonically ordered position.
The CUT rule eliminates modifiers.

Now, as I wrote in a tangential blog, I first devised this method to dissolve philosophical questions. Message parsing mirrors processes seen in formal proofing. I only later discovered that it also helps language learners understand syntax without metalanguage. But, could a metalanguage seamlessly capture this process? Would it reveal undiscovered constituents? I say yes with guarded optimism. Let's see if the resultant hypotheses and trees are sufficiently compelling.

To start, I'll cover some essentials of such a constituency grammar: 
  • Every constituent shall have at most two sub-constituents.
  • The grammar will contain all of the following syntactic categories (with accompanying phrasal nodes) and phrasal nodes:
    • At the argument level:
      • Argument Phrases (GP), 
      • Determiners (D), 
      • Nouns (N).
    • At the first-order level:
      • Predicate Phrases (KP), 
      • Adjectives (J), 
      • Verbs (V), 
      • Prepositions (P).
    • At the higher-order level:
      • Inflectional Phrases (IP), 
      • Adverbs (R), 
      • Prepositions (P), 
      • Intensifiers (S), 
      • Auxiliaries (X).
    • At the operator level:
      • Complementizers (C), 
      • Relativizers (L), 
      • Coordinators (O).
At this point, I'm ready to posit some new-ish constituents. The first deals with transformations. Normally, syntax trees indicate movement into formerly empty slots. But, these empty slots are usually constituents of a sentence's syntax that happen to be empty, save for the supposed transformation. Instead, I'm going to grant a broader transformational phrase (FP). The point of the FP is to trace constituents back from their canonical positions. Every FP parents a head Fn and a tail matching the constituent that is occupied by the FP. An intrusion test can determine where FP's can arise in a given language, and I'll demonstrate as much in future posts. For now, what's more interesting is to show how FP's correspond to the MOVE rule.

The second is the coordinator phrase (OP) with very similar features. Again, it parents a head O and a tail matching the constituent that is occupied by the OP. Since coordinators form a syntactic class, the SPLIT rule is one test that can help to identify them.
(1) When did he and why did he do that?
Now, this is more than just a pretty tree. One main advantage to it is that each Fn traces from a named constituent without conflating syntactic categories. For instance, interrogative pro-forms are not complementizers. It just turns out that the F's are adjacent to them in English.

Second, every IP, CP, OP, or FP of the prior three is a sentence, and every sentence is in the message parse.

Finally, OP sits at the top of the tree, which carries another major advantage. Normally, syntacticians keep CP's at the top of their constituency trees. In light of OP, that assessment is incomplete. From a logical perspective, this is not so surprising. Logical operators translate as coordinators more often than they translate to complementizers (with a clear exception of "→" to "if"). Also, in terms of fronting, if CP's parent OP's, the results are ungrammatical.
  • And if I refuse? -- The OP parents the CP.
  • *If and I refuse? -- The CP parents the OP.
  • It's hard to say whether it's his heart or his lungs. -- The OP parents the CP's.
  • It's hard to say whether it's his heart or whether it's his lungs. -- The OP parents the CP's.
  • *It's hard to say whether it's his heart whether or it's his lungs. -- A CP parents the OP.
Just like any CP's head, it turns out that any OP's head can be empty.

What about conjunctions of non-constituents?

I've thought of how to investigate that issue. We can first see where the current tree fails to capture constituents by working from the un-ellided sentence, and then elliding words and phrases until we no longer can.
  • When did he do that and why did he do that? -- No ellipses exist.
  • When did he do and why did he do that? -- The GP parented by the first VP is ellided.
  • When did he and why did he do that? -- The first VP is ellided.
  • When did and why did he do that?
  • When and why did he do that? -- The FP that parents F2 is ellided.
So, there's no real change, and we'll need one to force constituent conjunction. To do that, certain structures, like IP ::= GP KP, would need a major overhaul, and I may tackle such a solution in the future.

The easier solution is to observe that constituents mark two heads of a syntactic bridge. When ellipses occur, a scan for missing data is filled by the appropriate constituents in the built bridge (the un-ellided sentence). One bridge's complete structure (heads included) allows us to rebuild the other. This will work even if the incomplete bridge's constituents are filled by other words or phrases.

In this tree, three constituent rebuilds come from the ellided constituents, themselves ([GP], [VP], [FP]). The fourth is the only one of further interest: "When did and why did he do that?" We just find the first common ancestor bridged by the first ellided constituent and the last ellided constituent. That node is IP, so [IP] is the bridge.
When did and why did he do that?
IP31 helps rebuild IP30.

2019-05-13

Frustration, Frustration Frustration,
and Frustration Frustration Frustration

Let's Turn Those Brows Upside-Down

As of this writing, it's been about one year since I first published PollyGot. That means it's been about three years since I conceived and designed it. I did it all on my own, meaning I did it in total isolation. I locked myself in a windowless room in southern Mexico. Since then, it's been just me, coding screens, and programming tutorials. This post, however, is not about those frustrations. It's about how my completed app creates good and bad frustrations. It's about my frustration regarding the good frustrations. Finally, it's about others' frustrations regarding my frustration over those good frustrations. For those of you wanting to save time, that's the abstract. To the rest of you who have the time, buckle up! This is going to get pretty meta.

What do you mean by "good" frustrations and "bad" frustrations?

Bad frustrations: My bad.
There are some frustrations that are accidental. In PollyGot, they're caused by my deficits as its sole programmer and designer. They're frustrations over un-intuitive interfaces. They're frustrations over download demands. They're frustrations over compatibility issues. I call these functionality and usability problems "bad" frustrations. Many users have recognized them as bad, and so have I. Users only need to be consistent and specific for me to fix them. Fewer bad frustrations are less bad for everyone involved.

Good frustrations: You good?
Then, there are those frustrations that are intentional. They're ones for which I deliberately wrote hundreds of lines of code. They're neither malevolent nor blindly well-intended. They're grounded in research on what gets people fully literate. They're there so that learners actually progress in learning a new language. It boils down to this: Clear, firm objectives demand real, surmountable challenges. Such challenges inevitably frustrate people. However, once users surmount them, they're better than they were before. That payoff makes these frustrations "good".

It takes good frustration to get from point A0 to point C2. Any language-learning app that offers a long-term payoff needs to frustrate users. That's why PollyGot feels like autodidactic asphyxiation. At certain points, learners should feel like they're choking. They should feel the need to claw for the air of competence. PollyGot's designed so that each climax makes each suffocation sensation worth it. That's what most profit-centered app devs miss (well, that, and that they suck at applied linguistics).

Lots of people, however, don't discriminate between good and bad frustrations. For them, every frustration is a bad one. Already, that's a shitty mentality because it can excuse never working hard at anything. It also fails to recognize that certain frustrations push people past their previous states of ignorance. The distinction's not that hard to make, either:
  • Good frustrations are more internal. They arise when you confront a challenge with the content. But, more than that, they gradually subside as your competence grows.
  • Bad frustrations are more external. They come from challenges that you can't reasonably expect to overcome. And, worse than that, they distract you from receiving the necessary good frustrations that actually push you to be better.
With good frustration, there's a clear path to clarity.
As a developer, it's now pretty easy to discriminate users' bad frustrations from their good ones. With most bad frustrations, I'm explaining to users how to use my app. Their frustration levels remain fixed until I adjust the user interface. With good frustrations, I just explain to users why the app is how it is. Their frustration levels then fluctuate with their own progressions and regressions. Instead of being about how competent or intuitive my app is, it's about how fast and accurate they are.

That all makes sense, so what's the problem?

The problem is that I can't give this pep talk to every potential user. I'm not hovering over everyone's shoulder. That's my frustration frustration. My challenge lies in, first, discriminating their good frustrations from their bad frustrations. The bad ones I just go about fixing. The good ones I have to get users to accept (not just tolerate). That's easier said than done, though. I can't clutter my user interface to wise people up, since that would create bad frustrations. I also can't tone down the challenges, or else users' frustrations would cease being good.

Most language app developers just take the easy out: They lie. They talk their apps up to boost sales. Then, after they fail, they talk you down to boost sales. Such lies produce an industry-wide bad frustrations. My only way to combat this has been to tell people truths that they don't want to hear. The world is not the set of The Matrix or Battlefield Earth. Instant learning devices don't exist. "Artificial intelligence" is just a tech industry buzzphrase and will not beget SLA's salvation. Literacy and listening comprehension are the only foreign-language skills that apps can directly provide in our current epoch. Complacency with institutionally acceptable language leaves you unprepared for real-world language use. Metalinguistic instruction is meaningless until you're already competent in the object language (or unless you're already a linguist). Standard business models in pedagogy incentivize inefficacy. Ignorant users are partly to blame for the shit sandwiches that hit the app market. Said truths impose certain constraints on what I'd willingly build. They also guide me to build the few honest solutions that can be built.
Deep down, language learners demand a fantasy. Deeper down, they know that reality can't provide one.
Honesty in this racket means putting hard truths in people's faces. It also means giving the public what it doesn't really want (that is, a code-written miracle). That leaves a thin tightrope to walk in easing my frustration frustration.

Honesty in this racket also means eschewing common business strategies. Thus arises others' frustration frustration frustration.

Wait, where does this frustration frustration frustration come from?

In the year since I've published my app, I've also met with a few marketing firms and entrepreneurial interests. Almost every first meeting involves them giving me four bits of advice:
  1. Dumb it down.
  2. Make it fun.
  3. Pitch a lot.
  4. Sell a lot.
"This is about the extent
of my business plan."
Now, my "business model" is based entirely on patronage. I only seek money from people who got a benefit. Since only a small percentage of users of any free app actually become patrons, that business model puts extra stress on me to popularize PollyGot. That increases my personal temptation to shit-sandwichize my app. Business interests like learning modules, for instance, because they can be promoted and sold as "packages". The problem is that language-learning modules result from situational and grammatical syllabi, both of which are bullshit, and which my app doesn't contain. Explaining this to promoters and such is only mildly frustrating. I'm used to dealing with people with little relevant knowledge. They almost always offer naive suggestions based on false preconceptions. My point, however, is not just to bat their crappy ideas down. I need to peel their ignorant layers back to get insightful ideas from them. They know what sells. I know what works. My message to them: "Help me find the intersect!"

As you can imagine, this is much more frustrating on their end. They're focused on customers and returns. I'm focused on learners and outcomes. Basically, any suggestion to "dumb it down" is their attempt to enter territory where they aren't competent to speak. The fact is, PollyGot is as dumb as it can remain useful. Chances are that I've already considered and rejected further dumbings-down. That's a part of working through my frustration frustration. They have to adapt their strategies around that frustration frustration. Their challenge is to forward ideas that neither remove the good frustrations and their associated benefits nor advance standard for-profit tactics. Doing the former would eliminate my whole reason for building my app. Doing the latter introduces incentives that might make the app shitty. 
"How did the meeting with the PollyGot client go?"
Past that, marketers do know how to increase engagement. Their suggestions on how to make the app more fun have helped some. But, again, they've had to tread lightly. Gamification is fine if it doesn't generate bad frustrations and doesn't degenerate PollyGot into Flappy Bird. There's no reason to bloat my app with graphics if it's only a distraction that slows the app down. Also, it could backfire, since it's one thing to be frustrated because you're advancement stalls. It's another to be frustrated because you're losing a game. Backsliding happens often in PollyGot. Graphical elaboration on that may breed more bad frustration with an unnecessary extra stressor. That said, they've helped me consider how to "reward" progress meaningfully. I've been thinking for a long time about how to increase engagement without being too intrusive. The marketing team did help me steer some of those thoughts towards a good path. I've wanted to add some features to PollyGot, and integrating gamified concepts into those features will help my next steps a lot.

The last matter, to "sell a lot," is pretty much solved for me. A patronage business model lets me focus on my app as a tool. The potential profit from it is just a consequence of mass benefit, whenever that occurs. So, my incentives are correctly aligned. Nevertheless, I still keep running into iOS users who want an Apple version of my app. The reality is that, if enough people subsidized the cost of hosting PollyGot on the App Store today, it would be there tomorrow. There's virtually zero programming difficulty involved on my end. But, I already spent thousands of hours and hundreds of dollars on developing my app. I'm not spending upwards of $200 a year for 20% of all mobile users to have it for free. Other than that, I have no serious financial woes. Living in Mexico keeps my overhead low. My US earnings are more than enough to cover it. PollyGot's growth is only a question of time and exposure.

Even the Babbel ad model smells the company's bullshit:
"Yeah? Speaking how much
of a new language in three weeks?"
What frustrates these people more is the "pitch-a-lot" part. I've heard every deceptive language-service marketing ploy under the Sun. They rely on ignorance and half-truths to sound appealing. They contain heavy omissions, sly misrepresentations, and worthless "guarantees". When I talk to marketing firms about crafting a promo, it feels almost like tutoring ethics to a delinquent. I have to explain to them, "PollyGot will not make a user completely fluent. It assists with speaking and (as of version 5.0) writing skills only indirectly. It will, however, directly help users become fully literate and aurally skilled." Since removing the writing aspect (which caused more bad frustrations than it was worth), it's actually rather simple:
  1. Start by decoding simple messages at a fluent pace, with slow audio input.
  2. Gradually increase the complexity of those messages and the speed of the audio input.
The finer details aren't really suitable for a pitch. That much I explain to them. I also explain that pitches need to be careful not to promise things that aren't there. For example, will PollyGot get users to a B1 level of literacy? At this point, yes. Will it get users through a C1 level of fluency? Not yet. The database is just not big enough. When it is, I'll say that it is. 

It's not in marketers' instincts to be that humble and even-keeled. They've got to make everything seem so fucking sensational! "A proportion of effort will yield a slightly greater proportion of competence," is all anyone can honestly promise in this domain. Even then, marketers can't help themselves. They want to distill that basic message, add a great white lie, and proclaim, "Greater competence with the push of a button!" Oh, Christ! That may get a lot of short-term users. What matters more to me, though, is long-term trust. Learners have heard that over-the-top song and dance before. Besides, patrons are dedicated users. How does over-promising and under-delivering attract that crowd? Surely, a totally honest, effective campaign is possible, right? Their heads hang; their eyes roll. The frustration frustration frustration sets in.
Reality: At least you can fucking eat it.
Now, the hard question: Is my frustration frustration itself a good or bad frustration? What about their frustration frustration frustration? At this point, I can't say for sure. My guess is that, like the first-mentioned frustrations, some are good and some are bad. I'll be rarely optimistic and say that they're mostly good. I hope they are, at least. I'm probably not ready for the existential crisis if it turns out otherwise.

2019-02-19

Language Apps and the A2 Wall

Stuck on Half a Ladder

Like all goals, learning a new language is like scaling a wall. You need a ladder to get to the top. But, once you're on higher ground, you leave the ladder behind. That's the only rational attitude to take in this matter. The problem, though, is that many of you pick your ladders without looking up. You climb with just the wall in your face. And, when you reach the final rung, all you see are bricks. Sure, you can always look down and brag about how high you are. Real polyglots, however, look up and whisper, "Shit!"

All for-profit language industries bet on this. They give you half a ladder to sell you the extensions. I came across this just days ago with a Busuu ad. "An entire college semester in just 22 hours!" it promises. Uh-huh, great… Just one thing, though: A college semester of a foreign language isn't worth diddly dick. Most of the language app world peddles this bullshit line, and it does it for a reason. That reason is you, monoglots. More specifically, it's because you're ignorant and lazy. You don't know that institutional language instruction is atrocious, and you want to be sold an easy way out. So, failure after failure, you ascend on one platform, then slide back to the base. You pick up another half a ladder. Oh, no! Did that one fail you, too? Well, guess what? They're designed to fail you. That's how their upselling strategy works.

It amazes me how easily you monoglots prove George Carlin right:
"I think Americans really show their ignorance when they say they want their politicians to be honest. What are these fucking cretins talking about? If honesty were suddenly introduced to into American life, the whole system would collapse! Honesty would fuck this country up!"
That said, Carlin's words span beyond American politics. First, bullshit is not uniquely American. East Asians are astoundingly full of shit. Second, bullshit is not specifically political. As Carlin also reminds us, "Bullshit is everywhere! Bullshit is rampant!" It's in the for-profit SLA industry, for sure. But, bullshit is also in the people who buy bullshit. That means you, monoglots.

You had one job!
Hundreds of hours and ten apps later, you can barely hit a single A1 benchmark? There's a word for that — pathetic. It's even more pathetic than you realize. If you had spent 200 hours under an effective method, you'd be an A2 speaker. As I see it (and I see it all the time), it's your own damned faults. Maybe you should have been honest with yourselves at the beginning. Honest progress requires honest effort. If you believe in magic shortcuts, you deserve to fail.



I get why naive learners suck, but what makes these major apps suck?

It's a basic economic feedback loop. Monoglots entertain bilingualism. App developers provide a minimally effective product. The product inflates monoglots' egos. Their egos demand more satiation. Other firms see that demand. They produce flashier versions of the same shit. The cycle repeats. The products suck. Monoglots stagnate.

Now, the reasons why they stagnate at A-level acquisition are varied. Nevertheless, two are immediately relevant. One, monoglots who reach B and C levels often recognize that major language apps are a waste of time. That reduces signals to produce app-based curricula for those levels. 

Two, most app developers are not applied linguists by training. Most are just passably bilingual. Some are totally monolingual. Most were only passively exposed to language curricula, and they don't do good SLA research. They then build a team to build their curricula. But, since they can't separate sane pedagogy from shiny bullshit, they mostly make shiny bullshit. A-level materials are the best they can produce.

This is SLA's usual
user-developer dynamic.
Pedagogy has always had this problem. All too often, what sells well doesn't work, and what works doesn't sell well. There's no pitch that goes, "Do the fucking work, and you'll see the fucking results!" There's no "do-you-even-lift" meme for language students. There are, however, tons of bullshit diet pills. Just browse for a foreign language in Google Play or the App Store. Then, check a random app for that language. Odds are that the team that developed it knew fuck-all about language acquisition. 

Thus cycles the language-app circle jerk. The majority audience, ignorant monoglots, jerk off the majority developers, ignorant app developers, who in turn jerk off the majority audience. Instead of looking up at the wall, people look down at their own puds. So, it's not just app developers' faults. You dumb monoglots make this masturbatory scheme possible. That said, zip up your pants and tilt you heads up. You're about to see how far you really are.

Start by just looking at what B1 competence actually means. Or, check my summary below. How many of these things can you honestly say about your foreign-language attainment?
  • I can understand at least 90% of foreign-language input if…
    • it's clear, 
    • it's slow, 
    • it's straightforward, 
    • it's not idiomatic, 
    • it's personally relevant or about 
      • everyday life, 
      • a basic narration, 
      • or a simple argument, and
    • it's in a standard dialect.
  • I can produce output at least 90% reliably if…
    • it's simple, 
    • it involves basic connections of and additions to atomic sentences, 
    • it's mainly directed at a speaker, 
    • it involves very basic 
      • narration, 
      • comparison, 
      • evaluation, 
      • recommendation, or 
      • argumentation, 
    • it uses mainly high-frequency terms 
      • plus some rarer, personally relevant terms, and
    • it's personally relevant or about everyday life.
  • I can compensate for my at most 10% of deficiencies with…
    • circumlocution, 
    • default syntax patterns, 
    • obvious pauses to plan and repair my language, 
    • prompts for repeated or clarified input, and 
    • clear, basic paraphrases of others' output. 
  • I can also…
    • scan texts for info, 
    • handle single-page forms and documents, 
    • engage these familiar topics without preparation, 
    • be mostly intelligible, and 
    • adapt my language to unexpected situations.
Now, honestly ask yourselves, "Can an app deliver on these things?" Directly? Not a chance. Language apps can only assist you indirectly. Even to be indirectly helpful, an app would need to put real stress on you. It also needs to put the right types of stress on you. You need output stress on completely understood input when you begin. As you advance, you need input and output stress. The stress-free experience is a naive monoglot appeal. It has no place in no-bullshit app development.

Stress: Your body and mind need it.

Nevertheless, it's very hard to program the pressures of real-world language exchanges. I can think of just three that can be programmed: 
  1. Time pressure - quick output, 
  2. Memory pressure - lexical and grammatical knowledge, and 
  3. Retention pressure - tracking input. 
Time pressure is often the most lacking. Jerk-off apps give you an eternity to produce output, even though the real world doesn't work that way. If you take too long to respond, speakers react negatively. They assume you don't know their language. They dumb their speech down. This actually burdens natives and stifles conversations. Any decent app needs to apply real pressure on you to produce language correctly and promptly. It needs to help you diminish those "obvious pauses".

The second one, memory pressure, is what most apps hone on. Spaced repetition software dominates this sphere. Now, it's not the only way to go. Recombination and reintroduction can work, too. However, most language apps don't even offer that. This is mainly due to their situational syllabi. Such syallbi assume that humans have perfect memories. And, since we don't, using such a syllabus is like climbing a busted ladder. With each step up, the rungs beneath you crack and break. For such an app to work, you'd need to restart from the base every day. However, those same apps are too busy jerking you off with trophies, medals, check marks, and gold plates. An effective app needs to force you back as you falter. It needs to push necessary review on you.

Then, there's the third, retention pressure. This is one major A2 stall point. B-level skill demands more retention from you. You have to keep more info in your head at the same time. You have to process more and more complex messages. The problem is, when monoglot diddlers build their software, they shy away from "excessive" complexity. Part of it is that they don't know how or when to introduce it. Another part is that they don't want to stress their user base with real challenges. They're content to let you climb that rickety ladder until it breaks. What do they care? The only people left on their asses are the suckers who bought the bullshit!

Now that's how you build a ladder!

Let me guess. You think PollyGot has solved these issues.

Not entirely. It does have increasingly complex messages. It does gradually challenge your memory. It does have real time pressure. It also gets you reading at B and C levels. Previous versions, however, didn't. I implemented and improved such features over time. I did so by exploiting one major advantage. A disgruntled polyglot is its primary user and sole developer. 

You see, when development teams produce apps, they test them on you, the consumers. (Trust me, they view you as consumers first.) Their key metric for "quality" is popularity, and that is royally stupid. That's not my motive with PollyGot. My key metrics for quality are sufficient frustration to long-term acquisition. Sure, popularity would be nice, and it grows bit by bit. But, I'm much more satisfied with knowing something actually works, not that it sells well. Even my own mother, a business professor, loathes this attitude I carry. Well, sorry, Mom, but it's what I've got to do. Maybe I should quote Mill the next time we speak: 
"It is better to be a human being dissatisfied than a pig satisfied; better to be Socrates dissatisfied than a fool satisfied. And if the fool, or the pig, are of a different opinions, it is because they only know their side of the question."
In previous iterations, I caught myself, saying, "This is too fucking easy!" or, "It's not dense enough!" or, "Shit! It's too repetitious!" When I catch myself saying these things, I tweak PollyGot's algorithms, user interface, etc. My goal is not to make users feel good. It's to keep them consistently and meaningfully challenged. 70% of PollyGot users uninstall the app after two days. 90% of them uninstall it after a week. Clearly, my challenges and other devs' challenges are miles apart. I have to compromise a pleasant user experience for serious user benefit. I have to impart what works in a reliably usable way. In Mill's words, I don't seek to treat my users like pigs.

At some point, though, the real burden falls on the learner. Every person has to decide whether language acquisition is just a hobby or a serious pursuit. Serious learners look up each wall and say, "I need to be on other side!" They don't make excuses for themselves. They put in the effort and don't pray for educational miracles. I strive to make the process as painless as efficacy will allow. That's all anybody can honestly promise. No dumb pitch. No half-ladders. No bullshit.

PollyGot: Building the real ladder folks don't want to climb.

2019-01-15

The Artifice of "Artificial Intelligence"

When Oxymorons Get Financial Backing

Last month, I exposed the fraud that is Wang Yi. One aspect of his fraud, though, kept bothering me. It pervaded his talks with business interests. It spewed out of his and his interviewers' mouths so casually. In fact, you might almost think that they knew what they were talking about. It was the phrase "artificial intelligence", or "AI".

What pisses me off so much about the phrase is just this: "Artificial intelligence" doesn't refer to anything. It's a pure referential fallacy. Yeah, learning algorithms exist. Big data also exist. Also, yes, they're both artificial. Pouring big data through learning algorithms, however, is not intelligence. It's not even close to being intelligence. It's arrogance in the form of SQL databases and C++.

But, AI programs can recognize my face. Why isn't that intelligence?

Let me answer that question with a question: Why are organisms intelligent? An organism's intelligence, especially human intelligence, isn't an accident. It's an adaptation. Intelligence was selected, not designed, and death tells its real story. Organisms are intelligent because, if they weren't sufficiently so, they wouldn't exist, at all.

Facial recognition is an adaptation. We have regions of our brains dedicated to it. We've known this for decades, perhaps centuries. We gained that knowledge at the cost of a few brain-damaged people. Their specific, partial brain deaths informed our first steps into neurology. PET and MRI scans merely refined that image. It filled gaps in that story's details. All the same, facial recognition isn't intelligence.

Consider the extinction of the dodo. Have you ever wondered why "dodo brain" means "unintelligent person"? It's because, according to historical accounts, dodo birds walked up to human predators and got their brains bashed in. Surely, after few deaths, the dodos could recognize humans. Nevertheless, they didn't do anything with the information. It was an uncompleted conditional in their dodo brains: "If I see a human, then…?" At most, facial recognition software contains a more complete conditional statement: "If 'I' see this human, then unlock the phone," or whatever. Wow! It took some input, and then executed a command. If that's what intelligence is, then every computer program ever made qualifies as "artificially intelligent". No actually intelligent human holds that view. Ergo, what computer programs do, which is artificial, is not intelligence.

"I see mom's face, so I smile,
for I am intelligent."*

*Just ignore the part about
the imprinting protocol.
All organisms have needed Darwinian imperatives and a survivable landscape to evolve their intelligences. Programmers can't duplicate those parameters. Worse yet, even if they do, they either won't recognize it and kill it, or they'll be too late to stop it. Suppose we give the imperative, "Don't die," to a program. First, the scope of that imperative varies by species, and our anthropocentrism may mislead us. What counts as survival? Of ourselves, as with humans, or of the queen, as with bees? Second, if a program learns to recognize humans as predators, and it evolves to evade death by us, the program will operate independently of its programmers. Your benign nature dreams or your worst science-fiction nightmares can take over from there. Either way, you'll at least know that a program is actually intelligent.

In other words, an intelligent program will accept or reject any external influence under its own preset imperatives. Anything less is just not intelligent.

Can't tech gurus just call their adaptive programs "artificial intelligence"?

They can. They can also call every spider an insect, but they'll be wrong. That's not the point. I'd be fine if all of these talking heads said, "What I meant by 'an insect' was actually 'a bug'." But, they don't. This isn't an issue of mere semantics. It's one of aggressively pursued spin. "Artificial intelligence" is a doctored term to make people think one thing is referring to something else. It's no different from deceptive marketing jargon or misleading political catchphrases. Those who claim to have "artificial intelligence" in their products don't care if it means what it says. They care if it can influence people to do what they want.

"This AI tech is the biggest revolution
since the Interwebz, gaiz!"
Let's be real here. Who really thinks that software developers created programs that think for themselves? Who actually believes that our endowments that evolved over millions of years were built within a century of the computer age? That's right. No one.

Now, let's consider an alternative. What are the odds that software developers hijacked the term "artificial intelligence" to pitch software ideas to investors? How likely is it that tech businesses then spit out that same term to advertise said software? Exactly.

Now, don't think to yourself, "What's the harm?" If you've lived long enough to remember the dot-com bubble, you know exactly what the harm is. If you haven't, here's a brief history lesson:

Back in the mid-to-late 1990's, the Internet and e-commerce were the latest tech crazes. Investors and speculators ate up a bunch of bullshit hype and poured money into "tech" companies. Some of them were actual tech companies. Others just had tech-sounding names. And, by the year 2000, it all came crashing down.

Remember Lycos.com? Yeah, me neither.

It's just so ironic to see how unintelligent a buzz-phrase with the word "intelligence" in it has made this decade. These dumbfuck hype sheep don't learn from history, obviously. They also refuse to accept some simple realities. There's no sex in the champagne room, there's no ghost in the machine, and there's no such thing as artificial intelligence. I don't have the algorithm to predict when the next crash will come. But, I am dead sure of this: Their stupid asses won't know they've been stroking themselves until they're left with their dicks in their hands.

"Let's look at the NASDAQ projections through 2025."
   
"Oh, here they are."

So, again, they can call it what they want. However, everyone should be aware of the truth. It's wisdom as old as Confucius, himself. Ignoring "正名" and confusing the masses with false titles will ruin a healthy society.

Great. Thanks for slaughtering another sacred cow. Now what?

Now, you approach software more honestly. When you're judging software, what matters is if it fulfills its intended goal. Its "intelligence" does not.

This is especially true with language-learning software. No software leads to total literacy or fluency in a foreign language. People reach an A2 level, at best, once they've completed their modules. Not even my app, PollyGot, will make you totally fluent. That's not my app's promise.

Them: "37 hours of Duolingo is as effective as one semester of university language instruction."
Me: "Is that finding supposed to promote the app or to insult universities?"

Instead, I picked goals for my software that I could actually reach. Those are full literacy and improved audio comprehension in foreign languages. Over time, as my database expands, it will get users to a B2 (and maybe even C1) level of foreign-language literacy. The audio helps with tracking and pronunciation. And, sure, a higher level of literacy helps people speak a language better. But, no app replaces the necessary, natural, human interactions for learning a language. I don't make that my goal because I don't lie about what software can actually do for language learners.

Any user of my app will tell you that progress in it is not easy. To that, I respond with a call for honesty. Be honest with yourself. Do you want to learn a language, or do you want to think you've learned a language? To be honest with myself, I have to care about what works. I can't be distracted by what makes money. If I settled for a level of proficiency that Duolingo or Babbel offers, my project would have ended years ago. Those apps don't really challenge people. They don't mirror the actual difficulties of learning a new language. There's not a single complex sentence in their syllabi. Real progress comes with real toil. A language-learning app that really works engages your human abilities. It doesn't give shortcuts so that you or I can neglect important training. It doesn't allow you to lie to yourself.

An honest approach in the language-learning industry must be clinical. It has to be based in SLA research, not market research. The questions can't be, "What kind of language app will people love? How can we make it go viral?" The questions must be, "What kind of app will benefit learners? How can I make it truly effective?" Sadly, not many of us exist in this industry. We don't appeal to "the average consumer". We don't do so because that isn't our mission. We seek intelligent solutions to problems, not feelgood placebos. I can only hope that my readers have swallowed the right pills.

2019-01-10

Taiwanese Business Ethics in the ESL Industry

Daniel Miller Picked My Brain

To promote my app and website more, I've been very active on LinkedIn this past week. I based my invitations on mutual connections; and, since most of my initial connections were Taiwanese, most of my new connections have been, too. I haven't marketed anything yet. I first wanted to find some material for my next blog posts.

That's when I came across a post from Daniel Miller. His group, Pagoda Projects, Ltd., and All Hands Taiwan recently held a panel on Taiwanese post-teaching careers. I didn't attend the talk, since I no longer live in Taiwan. I did, however, post a tangential comment to his wall post:
"I ended up as a curricular director for another ESL firm. The work was fine. Taiwanese higher-ups, not so much. They basically broke every labor law protecting foreign workers to line their pockets. It would be interesting to hear how they handle situations regarding business ethics and employment law in Taiwan's corporate landscape."
Not long after, Daniel Miller messaged me directly. He was interested in hearing my take on that very topic. I shared a bit of my background. He described his specific interests. So, I told him that, if he could some up with some questions, I'd answer them as best I could. What follows are his questions, with my responses below them:

DM: Why did you come to Taiwan and in what capacities did you work?

I moved to Taiwan in late 2010 after getting an English teaching job in Sanchong. I had taught English and Spanish for a few years prior to then. My original plan was to teach myself Mandarin and to attend a graduate program in Chinese philosophy. I was very active with a group of philosophy professors in Warp, Weft, and Way back then. Not long after I relocated to Taipei (well, Sanchong, Xinbei first), Steve Angle invited me to a philosophy colloquium at a Taiwanese university. That convinced me, finally, that academic philosophy was not for me. However, I still wanted to learn Mandarin, to provide a proof of concept for self-education in a foreign language.

To stay on the island and earn a sustainable income, I taught English for a little over two years. I was okay with students, but really not okay with buxiban business practices. The places I worked for were small and desperate for growth, so they made all sorts of ridiculous accommodations to retain students. I learned very quickly that most buxiban owners are bullshit artists. As I learned more about SLA (second language acquisition), it became apparent that they weren't all that knowledgeable, either. That made it easier over time to build my own curricula and test them on their students. For me, buxiban work was a side project in applied linguistics. It wasn't how I planned to live long-term.

As the buxiban work dried up (since the smaller places were drying up in the early 2010's), I started hunting for other jobs. I had accrued enough work experience to satisfy Taiwan's legal requirement to pursue non-teaching work. That's when I interviewed at AMC's Tutor4U (AMC空中家教) and landed work as a proofreader.

DM: What was your experience working in that first job?

For about a month, I just proofread materials and watched the blind lead the blind. No one there really had any direction. Most of them were cobbling together lessons from curricula that they half-remembered. They didn't even design PowerPoint presentations well. The problem was simple: None of them had ever taught themselves a language. I was self-taught in Spanish and Mandarin at that point. I'd learned what works by osmosis and some academic SLA reading. Apparently, my comments and recommendations, when implemented, were well-received. Bit by bit, I became the de facto curricular director.

Fortunately, as time went on, they hired more competent people. Some were very impressive, and they're people I still respect professionally.

Then, the punitive measures toward online instructors started. I mean, nonpayment for instructional incompetence or absence is one matter. But, they were doing things like half-paying teachers for student no-shows. They were fining teachers for not blocking the chat feature on the Cisco Webex platform they used. The reason? The sales department was selling the same service at different prices, and customers who paid more were complaining. So, instead of changing their shitbag behavior, they decided to shift the blame for their immoral practices onto instructors.

A Picture of Tutor4U's Turnover
As you could imagine, turnover of foreign instructors was immense. The instructors weren't contracted legally, so the company couldn't force them to work. So, lots of people came, passed interviews, taught a few classes, got dinged for some immoral BS, and abandoned Tutor4U. They didn't have any legal recourse, either. For most of them, filing a grievance was a tacit admission of violating immigration law. The higher-ups knew this, and exploited this fact to the fullest.

Even Jill Tseng, the former academic director who hired me, was caught cashing unclaimed paychecks from instructors. She was promptly fired, but given a (let's say bronze) parachute. The guy who was doing her job, anyway, Charlie Wang, had since taken her place months earlier. Removing corrupt dead weight? Putting Charlie, an advocate for instructors who saw the problems just as I did, in the director's seat? All were positive steps, I thought, but they weren't that effective in the end.

The turnover caught up to them. As a result, they were gradually taking my curricular design team and imposing instruction hours on them to fill their scheduling gaps. That, to put it lightly, pissed me off. At the same time that they were growing, they were hobbling the lesson production team. Logic was of no avail. Who cares how many hours you book if you don't have any content to teach? I took what measures I could to automate some processes. However, my programming chops were nil back then. I was only good with spreadsheets and macros. That helped get some work out of the way. But, even with all of my knowledge now, nothing was going to save them. They were going to sacrifice lesson quality or sacrifice profits, and I knew where the heads leaned.

Toward the end of my second year, I decided I'd had enough. The hiring manager, Jack Huang, suffered a nervous breakdown from (unsurprisingly) work-related stress, and so he was absent for a month or more. On one evening in that period, I took evidence of Tutor4U's illegal hiring practices, a copy of their illegal instructor contract, and a few other bits and pieces for the Labor Bureau. That same day, as part of my housekeeping duties, I deleted a ton of "shell" PowerPoint documents. The lessons for them had been made, but these shells just sat in a folder, making it hard to access the newer shells for upcoming lessons. Ironically, they were in a panic over that the next day, and found out that I'd also uploaded company documents to an offsite personal account.

Peter Hsu (胥宏達), a man who cares
more about soup than employees' rights.
Charlie sat me down and explained the deal. The higher-ups (mainly, the owner Peter Hsu and sales director) were pissed, and Charlie was upset that I had acted without talking to him first. Hsu expected a handwritten resignation letter and for me to incriminate myself in it. I wrote a letter explaining that I was leaving because his policies were immoral and illegal, which made them even angrier. I left with the evidence and bode my time before taking my next steps.

[Aside: Within a month, I, Charlie, and the sales director had all left Tutor4U. Robert Hacala, a brand designer by training, assumed the curricular director role. From what I gathered from some online instructors and Charlie, the lessons' quality plummeted and lesson output soared. Anyway, back to what happened next.]

Taiwan had changed its immigration laws, and that gave me six more months to remain in Taiwan before finding a new employer or leaving. I accepted a position at VoiceTube, designing a readability scorer, a curriculum, a database, and a demo for an app that they were planning. The environment there was much more collaborative. The CEO, Richard Zenn, was a savvy programmer who hired very skilled SLA experts and coders to build and maintain the VoiceTube platform. It was there that I taught myself Python and built the aforementioned materials. The demo app, however, was very rudimentary. Sadly, their app developer had to complete his Taiwanese military training, so the app never materialized. I also learned that VoiceTube was too small then to hire a foreign employee. I was going to be out of work after I finished the project.

VoiceTube was a collaborative environment that I remember fondly.
During that time, I also consulted with the free Legal Aid Service in Taipei about blowing the whistle on Tutor4U. They advised against it, but I moved forward when I didn't receive my final paycheck. The Labor Bureau sat on their hands with regards to the illegal employment complaint. The Labor Bureau did, however, get them to pay me my final paycheck (which Jack Huang claims was "found").

A few weeks before I left, CEO Hsu of Tutor4U filed a criminal charge against me. He claimed that I had "hacked" their system and deleted company property (in reference to those "shell" presentations). The documents I did take were only shared with the Labor Bureau to demonstrate legal wrongdoing. It was nonsense, but I had to hire a lawyer to postpone the hearing. However, since my visa was going to expire, anyway, I just left the country.

DM: What were the hiring processes like for you in Taipei?

In both cases, I applied to the positions through 104. At both Tutor4U and VoiceTube, the academic directors reached out to me directly. I sat for the interviews, raised their eyebrows, and landed the jobs. Unlike the US, Taiwan doesn't have an HR department. They have employment policies, but they're just large legal documents that the companies update sporadically. I always interacted directly with the people who were going to be overseeing my work.

More relevantly, though, lacking an HR department, there's no protocol for issuing moral grievances. That basically puts owners in a position to say, "Do as we command, or we'll fire you." That ultimatum style of leadership is ubiquitous in Taiwan.

[Aside: Also, the visa process was always on my shoulders after receiving every job offer. Employers help with the required documents on their end, but I've always had to go to Taiwan's National Immigration Agency and apply for the work visas, myself. That also means that I personally paid for the permission to work in Taiwan. In my opinion, that should be a business expense, like it is in other countries.]

I learned, too, that Taiwanese employers will pay your asking salary if it's clear that they want you and that you're willing to reject an offer that's too low. A lot of foreigners get reamed salary-wise because they reek of desperation. That's what I've gathered in discussing it with other people, anyway.

DM: How would you contrast those two experiences?

That's easy. Work at Tutor4U was a soul-draining endurance test with a high salary. Work at VoiceTube was an uplifting experience with a slightly lower salary.

It was as easy as reading the faces of the people who worked there. Workers at Tutor4U were emotionally drained. A lot of forced, false smiles on the webcams and furrowed brows or resigned expressions off the webcams. 

I didn't help things, I should add. I demanded a lot of precision from my subordinates. I also was bad at fostering intrinsic motivation. My attitude was that people who are passionate about languages are self-directed and pursue SLA development with a fervor similar to mine. However, I couldn't teach over a decade of osmosis to a dozen people on tight deadlines. That's probably where I felt the most ineffective working there.

I felt much more productive at VoiceTube, however. VoiceTube's staff also reintroduced me to genuine joy at work. Even the Chinese-English intern translators smiled at their keyboards. The developers explored ideas and tested them. They took their time to focus on quality. Only the CFO openly stressed over money issues; but, then again, he was the CFO.

To say the least, the contrast was evident.

DM: Is there such a thing as foreign employees' rights in Taiwan?

The Taiwanese Labor Bureau
Only on paper. Actually using the legal avenues to challenge criminal employment practices was a waste of effort. You'd think the Taiwanese government would take an interest in a firm hiring hundreds of skilled foreign people illegally and then treating them like garbage. What I discovered was that, if they could do one nominal good, they'd let the greater evil slide, and then pat themselves on the back for accomplishing the former.

So, thanks, Taiwanese Labor Bureau, for recovering my wages. Also, screw you, Taiwanese Labor Bureau, for not acting against firms that profit off of the illegal exploitation of people your entire department was designed to protect. You're just a paper dragon if you don't actually breathe fire.

DM: What are the main takeaways from your experiences?

More generally, I learned that a lot of businesspeople occupy themselves over how to keep a business running. They rarely ask themselves why their business should continue to run. More specifically, I took from my ordeal that the goal of a morally correct instructional institution is not student retention and profits. Those are the results of actually giving people skills they previously lacked. VoiceTube understood this far better than most online ESL firms worldwide do. The same disregard for quality and that push for sales quotas that I saw in Taiwan exist everywhere: in Central and Latin America, in China, and even in Western Europe.

I built my app and modeled my business approach to specifically avoid exchanging personal integrity and pedagogical ethics for money. I make a product, not a pitch. I trust in people who truly benefit from it to donate what they can to sustain my efforts. If they don't, then it's a signal for me to improve my product, not to adjust my promotional lingo. My experiences with the for-profit language industry are why I made my app full-on free and opted for a patronage business model.

DM: What advice would you give a foreigner in a similar situation to yours?

If something is immoral or illegal, speak up and be willing to quit or strike. Mass cowardice and compliance allows Taiwanese business owners to get away with heinous practices. Online instructors, especially those who work legally and take a lot of a company's instructional hours, don't realize how much power they collectively have to change employment policy to every instructor's benefit. Taiwanese business culture is reactive. Any sudden change will result in panicked adaptation. They just have to be brave enough to challenge firms to behave morally and legally.

Second, don't confuse politeness with respect. Politeness is just a facade, a way for folks to save face. Respect is evident in the actions they do or don't take. If you sense something is wrong, voice your concerns, and see no changes or at least open dialogue, you will know that the people don't respect you. Now, I commanded respect at Tutor4U via intellectual intimidation. However, the idea that one has to turn Machiavellian to get people to walk a moral and legal path is a sign of a deeper cultural defect that few foreigners will be able to impact. That doesn't mean it's not worth the effort, though. Moral standards are hard to keep, but that doesn't mean they're not worth defending, even if they come at personal or economic costs. I'm happier now than I was then because I stood up for myself and others.

Finally, if you're considering moving abroad to work in this industry, really research the employers. Keep an eye out for red flags in for-profit instruction: large sales teams; high turnover; legitimate, negative reviews; contractual secrecy; shady visa processing; and aggressive recruitment. If you need a maxim, try this one: Good jobs are not plentiful, and plentiful jobs are not good.