Language Log

Kai-Fu Lee has an interesting opinion piece in yesterday's NYT: –"The Real Threat of Artificial Intelligence":

What worries you about the coming world of artificial intelligence?

Too often the answer to this question resembles the plot of a sci-fi thriller. People worry that developments in A.I. will bring about the “singularity” — that point in history when A.I. surpasses human intelligence, leading to an unimaginable revolution in human affairs. Or they wonder whether instead of our controlling artificial intelligence, it will control us, turning us, in effect, into cyborgs.

These are interesting issues to contemplate, but they are not pressing. They concern situations that may not arise for hundreds of years, if ever. […]

This doesn’t mean we have nothing to worry about. On the contrary, the A.I. products that now exist are improving faster than most people realize and promise to radically transform our world, not always for the better. They are only tools, not a competing form of intelligence. But they will reshape what work means and how wealth is created, leading to unprecedented economic inequalities and even altering the global balance of power.

As I've observed several times over the years, automatic speech recognition is getting better and better, to the point where some experts can plausibly advance claims of "achieving human parity". It's not hard to create material where humans still win, but in a lot of ordinary-life recordings, the machines do an excellent job.

Just like human listeners, computer ASR algorithms combine "bottom-up" information about the audio with "top-down" information about the context — both the local word-sequence context and various layers of broader context. In general, the machines are more dependent than humans are on the top-down information, in the sense that their performance on (even carefully-pronounced) jabberwocky or word salad is generally rather poor.

But recently I've been noting some cases where an ASR system unexpectedly fails to take account of what seem like some obvious local word-sequence likelihoods. To check my impression that such events are fairly common, I picked a random youtube video from YouTube's welcome page — Bill Maher's 6/23/2017 monologue — and fetched the "auto-generated" closed captions.

Adrienne LaFrance, "What an AI's Non-Human Language Actually Looks Like", The Atlantic 6/20/2017:

Something unexpected happened recently at the Facebook Artificial Intelligence Research lab. Researchers who had been training bots to negotiate with one another realized that the bots, left to their own devices, started communicating in a non-human language.  […]

Towards the end of April, I gave a short presentation at the Penn Science Café  in a session on "The past, present, and future of AI". I mentioned this in a comment on an xkcd cartoon in "Machine Learning", 5/17/2017, where I also reproduced my opening Science Café slide:

Over the weekend, Fernando Pereira posted a wonderful account of these three eras, with some thoughts about the nature of the underlying problems and possible directions for the future: "A (computational) linguistic farce in three acts", Earning My Turns 6/10/2017.

A Quinnipiac Poll released on May 10 asked respondents "What is the first word that comes to mind when you think of Donald Trump?"  46 words were used by 5 or more respondents. The full list, with the number of responses for each word, is here — the top 15 words were:

idiot         39
incompetent   31
liar          30
leader        25
unqualified   25
president     22
strong        21
businessman   18
ignorant      16
egotistical   15
asshole       13
stupid        13
arrogant      12
trying        12
bully         11

For other reasons, I've recently been gathering word-linked information about features like frequency, concreteness, positive vs. negative valence, etc. So I thought it would be interesting to look at the (obviously bimodal) distribution of positivity found in this list, and perhaps the distributions of some more subtle properties as well.

From Simon King:

We need your help with the Blizzard Challenge listening test for 2017.

Please take part yourself, and encourage your colleagues and students too. Feel free to forward this message to your local mailing lists.

Speech Experts (you decide if you are one!) take part here.

Everyone else, do it here.

It's a fun test – you get to listen to paragraphs from children's stories! It takes about 45 minutes to complete; you can take a break at any point, then continue where you left off.

Deadline for completion: 15th June 2017

Today's xkcd:

Mouseover title: "The top search for every state is PORN, except Florida, where it's SEX PORN."

And the lesson doesn't just apply to maps, of course…

In a comment on "Political Sound and Silence II", 5/30/2017, and referencing "Trends in presidential pitch II", 5/21/2017, unekdoud asked

Are there are any trends over the Weekly Addresses for these measures? In particular, is speech duration or speech % correlated with median pitch?

There's certainly a relationship (r=0.55) between speaking rate (words per minute counting speech regions only) and median f0, in the Weekly Addresses for Donald Trump's first few months as president:

In "Political Sound and Silence", 2/8/2016, I compared the joint distribution of speech segment durations and (immediately following) silence segment durations in the Weekly Addresses of Presidents George W. Bush and Barack Obama:

Today I thought I'd add a similar graph for President Donald Trump's Weekly Addresses so far in 2017:

Or not. Andrew Mayeda, "World Bank's Star Economist Is Sidelined in War Over Words", Bloomberg 5/25/2017:

The World Bank’s chief economist has been stripped of his management duties after researchers rebelled against his efforts to make them communicate more clearly, including curbs on the written use of “and.” […]

A study by Stanford University’s Literary Lab in 2015 found the bank’s use of language has become more “codified, self-referential, and detached from everyday language” since the bank’s board of governors held their inaugural meeting in 1946. The study coined the term “Bankspeak,” a vague “technical code” that symbolized the lender’s organizational drift.

In an email to staff obtained by Bloomberg, Romer argued the World Development Report, one of the bank’s flagship publications, “has to be narrow to penetrate deeply,” comparing his vision for the report to a knife. “To drive home the importance of focus, I’ve told the authors that I will not clear the final report if the frequency of ‘and’ exceeds 2.6 percent,” said Romer, citing the percentage of the word’s use in World Bank documents analyzed as part of the “Bankspeak” report.

I've been downloading the audio for Donald Trump's Weekly Addresses from whitehouse.gov, as I did for George W. Bush and Barack Obama. And as I did for the previous presidents, I listen to the results and sometimes do simple acoustic-phonetic analyses — see e.g. "Raising his voice", 10/8/2011; "Political sound and silence", 2/8/2016. Recently I thought I noticed a significant change in Mr. Trump's pitch range, and a quick check confirmed this impression.

Early one evening last week, I was feeling sleepy, and said so. And a little later, I said "OK, I'm cashing in my threat to take a nap", and went into my bedroom to do so.

As usual, I took my cell phone out of my pocket and plugged it in to charge, which made the screen light up. On it I saw this: