Three Layers of Understanding a Message: A common thread connecting Astrology, Cryptic Crosswords, Racial Understanding and Brittle AI

The Bottle

John was strolling down the beach when he spotted a glass bottle washed ashore. Curious, he picked it up and saw that it was empty but for a piece of paper with some typed message. When he looked closer at the message, he was disappointed. It wasn’t in English. It looked Russian. Luckily, he had a Russian-speaking friend. “It is not Russian,” the friend said, “but Ukrainian. I think I can manage to make sense of it”. The message was from the magazine “National Geographic World”, part of a study to map out ocean currents, and had instructions on where to send information about where and when the bottle was found.

Three Levels of Messages

Three distinct phases can be discerned in going from spotting the bottle on the beach to understanding the message, and failure in any of those stages would lead to not understanding the message.

I here use the terminology from Douglas Hofstadter’s “Gödel, Escher, Bach”, where he uses the three terms Frame message, outer message, and inner message.

The frame message is simply the recognition that there is a message here. Had John ignored the bottle as just another piece of trash littering the beach, the understanding process would grind to a halt without even starting.

The outer message concerns an identification of the language of the message and identification of the mechanisms of “reading” that message. If John had not realized this looked vaguely Russian, and had no means of finding out, or no friend who could help read it, the process of understanding would also have failed. Here in this story are two layers of outer message understanding: language guessing and getting the friend’s help.

The inner message is what one thinks of as “the” message (here, the instructions of where to send information). But we must recognize that the frame and outer messages must be deciphered before this final stage can happen.

Analyzing situations through this lens

It is fun and illuminating to peer through this level-of-message lens at the process of understanding — understanding both how we read text and understanding how we “read the world”. In the rest of this post, I will look at Astrology, Cryptic crosswords, conspiracy theories, racial misunderstandings, and brittle AI systems.

I will limit myself to glitches that happen in reading either the frame message or the outer message. The first four examples below concern frame-message glitches and the final two concern outer-message glitches.

Learning to Program

When I was a graduate student, I assisted in programming labs where young undergraduates were learning to code. When they got stuck on something, they would raise their hand and I’d walk over to them and try to help. One time when I was summoned, they were trying to compile a program and were running into a syntax error that the compiler was helpfully clearly spelling out: they had not declared the variable they were using. I could not understand what they were trying to ask me given that the error was perfectly pinpointed by the compiler. Here is what we discovered after several minutes of talking past one another: the student did not know that they were supposed to read the compiler’s response. They just thought of that as random noise and were only looking at compiled or not compiled. They had completely failed to read the frame message “I am a message”.

Astrology, Phrenology and Stock Market “Technical Analysis”

Astrology seeks to understand the future via planet positions. Phrenology seeks to measure intelligence by skull shape. Both are led astray at the earliest phase of message understanding, misreading the “frame message”. They see the position of the stars as the bottle with the message and proceed to decipher it. The problem is that there is no message to be read, but “Seek and you shall find”: there are apparent patterns even in purely random dots. They read a non-existent frame message.

I wonder if predicting the stock market via “technical analysis” and discovering patterns such as Head And Shoulders falls prey to the same strong belief in the existence of a message leading to the inevitable “discovery” of patterns.

Cryptic Crosswords

For the uninitiated, Cryptic crosswords can be frustratingly hard. For the record, I myself am very near the “uninitiated” end of the spectrum.

The example here comes from a nice-looking guide to solving. Here is a clue for a 7 letter word: It turns out that in cryptic crosswords, a word such as ruined or destroyed or rearranged is a frame message and outer message rolled into one: it often means that the prior word is an anagram for the answer. Thus, this is a recipe for identifying where the message is (prior word) and what language it is (a shuffled version of the answer). Here, we are looking for an anagram of that means , and the answer is

Conspiracy Theories

Whereas the programming example above concerned missing the frame message and astrology concerned manufacturing a non-existent frame message, conspiracy theories traffic in hidden frame messages, where apart from the “obvious” message that most people see, there is a parallel hidden message. This is analogous to saying “Yes, there is a message in the bottle, and although benign, it is a mere decoy. The shape of the bottle is itself the more important message that only those in the know will understand”. For instance, the message James Comey tweeted about his dog’s death was, according to some in QAnon, a coded message about taking out George H. W. Bush. And here is an example-rich article concerning conspiracies about hidden-message in popular songs.

We now turn to glitches in outer message understanding.

AI Brittleness: Syntactic Analysis

When I worked at Google, my work took as input sentences and their syntactic analysis. A sentence such as “I ate fried rice” is parsed into its parts-of-speech and the role of the words in the sentence. Thus, is identified as an adjective, as the verb, as a noun, and further their relationships are identified: modifies , and is the grammatical object of the verb .

In English, the subject of the verb typically comes before and the object of the verb typically comes after the verb. A typical sentence is The persons arrested are seen after the verb.

But perhaps I was too hasty: I should have said “In English running text”. In newspaper headlines, by contrast, the grammar is not the same. One sees, for instance: . Here, the person arrested is seen the verb. In this sense, newspaper headlines are a different language (or, if you prefer, a different dialect). The system would parse as the arrester, not the arrested.

The syntactic analyzer made errors aplenty. In the framework of the analysis we are attempting, these were errors in understanding the outer message, i.e., in identifying the language of the message. Messages were parsed as if they were running English text, not as the choppy headline text they often were. The errors got hilarious when the “real” language of the message was even more distant, such as math equations. Many English documents naturally have equations in them, but the syntactic analyzer never paid any attention to the outer message and parsed these as English sentences. Thus, “2 + 3 = 5” might get parsed as 3 being the verb whose object is “= 5” where “=“ is the adjective. I kid you not: we did see these errors in every math equation.

Another delicious parse error was “Chicken Fried Rice” (which comes from the English dialect “Menu Entries”) parsed with as the subject and as the object of the verb . Imagine a giant chicken, complete with an apron and a chef’s hat, merrily tossing some rice.

The final parse error I will mention is skynetesque: “end user agreement”, with as the object of the verb

Armed with these examples of different language flavors leading to errors when confused with each other (despite all being purely English), let’s move to murkier waters where I must tread with care.

Cultural and Racial Misunderstandings

Let’s consider a single phrase uttered by two different people. I contend that that phrase means two different things when said by these two people. The phrase comes from different languages (although both English), differing in the cultural contexts of the speakers. Just as the grammar of running text can be different from that of headlines or menu entries, understanding one as if it was the other leads to misreadings. This is thus a failure of reading the outer message: the incorrect interpreter is picked. This is not unlike someone on Twitter not realizing they are reading a joke and responding uffishly (been there, done that).

The first hypothetical person I’d like to consider is a follower of Jainism newly-arrived in the United States. The wikipedia entry for Jainism summarizes this person’s food habits well:

The second hypothetical person is a prototypical white supremacist.

The phrase in question: “All lives matter”. It does not mean the same thing when uttered by these two individuals. One literally means “all lives matter” and is not said in contrast to “Black lives matter”, but the same cannot be said of the utterance by the second person.

In online discourse, we usually don’t know who is who. Tempers get hot, and charitable readings of sentences become impossible. What I said applies not just to this sentence but to many other polarizing sentences and phrasing.

Alas, nowadays we only have time to hear what the words mean and not what the speakers mean by those words.

I do unsupervised concept discovery at Pinterest (and previously at Google). Twitter: @amahabal