a3nm's blog

Figuring out adversarially that someone understands a given language

— updated

This post is about the following hypothetical scenario: Alice is suspected to be a spy from a foreign country, and so to be fluent in the language spoken in that country, called language L. Alice claims to be an innocent citizen with no knowledge whatsoever of language L. Of course, this is just what a spy would say!

You are in charge of interrogating Alice, and to figure out whether she is fluent in L or not. How do you do it?

Of course, you may be lucky and Alice could inadvertently out herself as an L speaker. For instance, she may use an L word accidentally1 (e.g., cursing, or counting out loud2, etc.3). Your task is also very easy if Alice has a thick L accent in the language that you share with her. But what is Alice is perfectly prepared, perfectly fluent in your language, and you cannot count on her to make a spontaneous mistake? Of course we want to avoid mistreating Alice (she might be innocent), and we can assume that she will reasonably cooperate with the interrogator (like an innocently accused citizen who wants to prove their innocence).

Here are some ideas of solutions to this problem: thanks to louis, Ted, olasd, and Tito for contributing some. There are also some solutions here which are already mentioned in a TVtropes entry with a wider scope)4; but maybe there are other possibilities.

Brain imaging

With access to enough technology, there is a solution which is probably foolproof and requires no cooperation from Alice. Just put her in a neuroimaging machine, and have her listen to spoken recordings in language L. By comparing with a language that she knows, and with a language that she doesn't, it should be possible to detect whether her brain is making sense of what she hears or not. (I'm no expert in neurosciences though, so I cannot promise this would work.)

In what follows, I consider this to be cheating: I assume we don't have out-of-band access to Alice's brain, and that we must test her via normal sensory interfaces.

The Stroop test

The Stroop test is the following task: you are presented with a sequence of color names which are themselves written in a different color (e.g., it could contain the word "blue" but written in red), and you must read the sequence of the colors in which the words are written (not read the words themselves). This turns out to be difficult, with people often messing up and reading the words instead of naming the colors. But of course the Stroop effect only works if the person knows the language in which the words are written.

So if Alice's performance on this task is different between color words in language L and nonsense words, then you can find out that she is familiar with language L. The Stroop test may in fact have been used for this purpose historically, as discussed in this skeptics.SE question.

Alice may be able to evade detection by deliberately slowing down on nonsense words to try to match her performance on L words -- and in particular to make sure that she never ever translates an L word, which would be a dead giveaway. However, I would assume that maintaining exactly the same performance (and same error profile) between genuine mistakes (on nonsense words) and contrived mistakes (on words of L) would be very challenging for Alice. Especially if you collect precise timings on her performance that she herself does not have access to.

Timing attacks

A generalization of this idea is to measure Alice's performance on other tasks involving both L words and nonsense words, and seeing if her performance is different across both classes. Like, test Alice's ability to memorize short phrases, some of which are made of nonsense words and some of which are well-formed sentences in L. Of course if Alice is fluent in L the second task would be far easier for her.

I guess one can come up with similar tasks, e.g., if L uses an ideogram writing system, I would expect that it would be easier, say, to find differences between two figures, or find occurrences of a figure in another figure, or other such tasks, whenever the figures used in the task are genuine L ideograms as opposed to similar-looking but nonsense ideograms.

This is a bit like a timing attack in computer security. It also resembles a bit the implicit association test, though I find it difficult to adapt this specific test to a task for Alice that would also make sense if Alice has no knowledge of L whatsoever.

Following verbal instructions

Another idea is to have Alice play a game where she must follow instructions as quickly as possible. The catch is that some of the instructions are nonsense instructions, and some are instructions in language L. If Alice spontaneously reacts to one of the language L instructions, then she is unmasked. Playing this correctly if you understand language L is a bit like a Simons Says game, whereas of course it is not especially difficult if you do not know language L. Of course, Alice may again be able to avoid detection by deliberately slowing down.

A similar idea is mentioned in the TVtropes entry mentioned above, about a possibly apocryphal practice by the British of shouting "Achtung!" to identify German spies. Of course variations of this idea can work if you are not overtly interrogating Alice but watching her without her knowledge, which I would also consider as cheating.

A related idea: inserting words from L in conversation, using them as loanwords, and seeing in conversation whether Alice understands one of them (i.e., forgot to pretend she didn't). A related strategy (asking a question in French in the middle of an interview in Spanish) was used by Ladislas de Hoyos to identify Klaus Barbie while he posed as the non-French-speaking Klaus Altmann.

Watermarked language

Another idea hidden in plain sight: how about giving some classes to Alice where you teach her language L, and see how she performs? If she doesn't react like a real beginner would, in particular if she uses just one word that you hadn't already taught her, then she is unmasked.

There is a meaner variation on this idea, but which requires much more preparation. You could design and teach Alice a constructed language L', which is very similar to L except that it is "watermarked" in many small ways that are difficult to remember5, e.g., the orthography of some words is subtly different, some words have been exchanged, etc. If Alice is a spy, you would expect her to mess up at least some of the time with errors influenced by L. By contrast, if Alice is innocent, the errors she makes would not be correlated to L.

This is not a very practical solution, and also I don't know whether it would work in practice. Still, if it does, I find it interesting that knowing something (L) may be a handicap in properly learning something else (L').

An explicit solution

To finish with an entirely different idea, another way is watch for a physiological response: read some erotic literature in language L to Alice and see if she becomes excited. Ironically, this very low-tech solution comes from a science fiction short story: I'm in Marsport Without Hilda, by Isaac Asimov.

Other involuntary reactions to language could also work, e.g., laughter (e.g., with jokes), disgust (e.g., with gruesome descriptions), anger (e.g., with insults), etc.

  1. There is a scene like this in the movie Inglourious Basterds where an undercover British agent is unmasked because he uses the wrong hand gesture to order beer. 

  2. Somehow I didn't find a standard name for the observation that even fluent speakers of a foreign language will often spontaneously revert to their native language when they are counting out loud. Yet, this seems obvious to me based on personal experience (from myself and others), and it's not hard to find people discussing this on the Web. 

  3. One example from my own native language: in French, the word "enfin" (in this context pronounced "'fin" /fɛ̃/) can be used (among other things) to take back something you just said, as a kind of verbal backspace key. Like "I mean" in the following utterance: "He looked pretty happy, I mean, not super happy, but...". From personal experience, some native French speakers, even when fluent in English, may mess up and use "enfin" in the middle of an English sentence. 

  4. I found that webpage because it is (up to duplicates) the only occurrence found by Google Search of the terms "Stroop test" and "Marsport without Hilda" (which I mention in my proposed solutions). This suggests to me that the problem I discuss here doesn't seem to have been widely addressed elsewhere. 

  5. A bit like trap streets in maps. 

comments welcome at a3nm<REMOVETHIS>@a3nm.net