Disappointment and Self-Delusion
and I it gets in his earlier award of what does it say
if he gets in his earlier a word of what this
it’s in an earlier old word of what is she
in an earlier old world and what is he
earlier are more important than what
a study of physics is more important is what
The study of Physics is more and more disappointing
are you so you are so few
are you so if you argue so if you
you and myself and if you
yeah in myself and in
in my self-delusion
the in in my self-delusion
in modern physics is self-delusion
modern physics is self-delusion hanging in
Modern Physics is self-delusion
October 30, text transformed by Automated Speech Recognition Failures: method 907d701b-6499-4145-8cb2-3a4f95363109 (telephone game solitaire with ASR)
This guy Gang Lu, who got his Physics PhD in a bad year for the industry, was pretty upset about how his life was going. So he got a gun and shot his adviser, he shot a couple other faculty members, shot a grad student, shot the assistant vice-president for academic affairs, and shot the assistant VP’s secretary, then shot himself in the head as the cops came to take him out. The movie “Dark Matter” (loosely based on the incident) was on hulu a while back, it’s not bad.
A couple of phrases from the guy’s suicide note:
- “The study of Physics is more and more disappointing”
- “Modern Physics is self-delusion”
The rest of the suicide note was suppressed. It was probably just poorly-reasoned rants. Anyway, I used method 907d701b-6499-4145-8cb2-3a4f95363109 (telephone game solitaire with ASR) using Dragon NaturallySpeaking 9 on a Windows XP box to generate more lines of text, then reversed the order with the initial text last and removed some redundant lines. Interestingly, even though I was using an old 1980s-era Sony Walkman mic, the ASR quality was pretty good; I had to put the mic farther away from my head and throw some 2Pac on in the background to get poor quality.
But what’s really interesting is the language models involved. Contemporary ASR uses acoustic models (which represent how a given speaker pronounces a given phoneme) and language models (which represent the types of words that a given speaker might use.) I didn’t train the acoustic or language models to my voice and document, so they were “generic.”
So: the new lines of this poem were generated
- by errors in the ASR’s understanding of generic acoustic model (i.e. how the ASR software developers believe their customers speak to the world)
- by errors in the ASR’s own generic language models (i.e. how the ASR software developers believe their customers uses language)
- based on a set of initial lines (i.e. the language used by a Chinese-native-speaking Physics researcher at the end of his life.)
- in search of a set of text that will be forever hidden (i.e. the language used by that Physics researcher in the full death note that has been suppressed.)
I think what I need to do is understand the connections between each of those language models, down to a formal leval equivalent to the mathematics used in the acoustic and language models of 1-2 and the grammatical level possible with 3-4. I think there’s a more fundamental question of whether the acoustic, language, and grammatical models are sufficient to represent the connections in any way that is interesting to me, because I don’t really know how to computationally represent all the other language models in the picture, such as of the author, the reader, other members of language-use communities that work together to make meaningful language use possible, … uh… man…
If I really wanted to find out that “hidden text”, I guess I’d find a corpus of complaints made by native-Chinese-speakers, determine what type of sentences followed sentences line the seed texts above, and use that to generate.
If you’re gonna be disappointed about anything, let it be about your own ignorance to understand the hard research questions that are there forever just out of your grasp. And if you’re going to be self-deluded, let it be about the possibility of your solving them. HA!