Preliminary Notes on the Concept of Mutual Language, pt. 1
Some semi-ordered conceptual background
Over the course of the coming {span of time I will spend publishing here}, I’m going to talk a lot about the concept of a mutual language and its importance in several fields, including artificial intelligence and broader human-computer interaction.
Because I myself seek to establish a mutual language with my audience (meta!) and have little patience for repeating the basics every time I write something on the topic, this will serve as our anchor point.1
In short, this post’s function and its features are inextricably linked ;)
A brief definition for “mutual language”
A mutual language is a semantic structure agreed upon and used by multiple parties, frequently within a closed community setting. In functional terms, mutual language is the thing that allows you to communicate effectively and differently with your partner, coworkers, bandmates, etc.2
Though many like to think of language as having some “inherent meaning,” the same sentence functions differently depending on its social context. Much like how a word is defined in context by the company it keeps, a sentence, paragraph, or document finds its meaning only in a world that extends beyond its boundaries.3 Furthermore, it functions differently depending on its context, and on the person on the other end. Language is a compression algorithm for an idea; mutual language is a system of minimally lossy decompression.
Extension of the concept to Human-Computer Interaction
TL;DR: Any two people must establish a mutual language if they wish to communicate effectively. Why should human-computer interaction be any different?
You’d be hard-pressed to find a person these days who hasn’t interacted with some piece of technology that meets both of the following criteria:
A key component of its functionality is to “understand” human input; and
Its ability to do so requires the human user to put a great deal of effort into helping it.
In other words, the human user specifically alters their input so that it will be understood “properly” by the computer, or finds some way to satisfice.4 Here, if a mutual language is established at all (i.e. the user figures out a way to express their query such that it yields the desired results), it finds itself skewed largely in favor of the computer’s baseline, as the computer is not a syntactically flexible agent.
Consider, instead, the possibility of a mutual language skewed human -- a language based primarily upon the natural linguistic patterns of an individual user. Though potentially imperfect -- by which I mean that the human user must still adapt somewhat to be understood -- such a model increases the likelihood that any chosen query has a minimally lossy representation in the mutual language.
An input that loses information in translation between man and machine loses less the more it can avoid being warped over the course of that translation. This follows naturally from the reduced distance between the endpoints of the translation, consequently bringing edge cases closer to the center.
We call this retention expressivity. A hypothetical system that establishes a more human mutual language with its user is more expressive along various dimensions.
A note on “humanity”
Prioritizing a more “human” mutual language favors a more “human” experience. As meaty i/o machines, our inputs beget our outcomes. The inputs that we reinforce through action -- speaking or typing words, physically behaving in various ways, etc. -- influence us most of all. We become the things we embody, etc. etc.
What we’ve got today feels like CGI in the early 2000s: it’s blowing people’s minds, but at some point in the maybe not-so-distant future we’ll look back at this and say, “Man, that tech is hilariously crude. I can’t believe we thought that was state-of-the-art. Making people do work to make the computer understand them properly? Ha!” And their computer will laugh along with them, having understood the humor of the situation.
Extension to the realm of education
To reiterate, the best way to communicate a thing to someone (human or otherwise) is to get as semantically close to one another as possible before spitting your thing into the translation tube. Think of this like whispering to your friend sitting next to you, rather than screaming to them from the top of the Empire State Building. The transmission requires less volume (energy), and the chance that something will get in the way of them hearing you properly is much smaller (no birds, planes, or supermen).
Taking a functional approach to language allows us to decrease the semantic distance to our conversational partners, as the only requirement for mutual understanding is the ability to identify a functionally equivalent framework that is already within their lexicon.
Take someone who doesn’t “get” math but is interested in military strategy. Explaining concepts in graph theory becomes nearly trivial, assuming you yourself are able to see the links between graph theory and military strategy. Then, introducing the symbolic notation becomes simply a convenient shorthand for talking about something they already understand, rather than the insurmountable barrier it otherwise would.
The concept of teaching through analogy is not by any means new. Any decent instructor working closely with a few students is well acquainted with the practice. In larger classrooms, however, this becomes impractical. Jane cares about military strategy, Bob cares about Pokémon, Joe cares about trains, and each of the three has no interest in the topics that excite the other two. What functional equivalent works for all three? And, so, we revert to teaching math through its notation.
An alternative: ask the class, “What is something that might require you to X?” Listen to their answers.
Backwards compatibility in education
I would be remiss not to touch upon the issue of backwards compatibility. Most people just don’t have it. Once they’ve grokked the thing that required grokking, they mysteriously lose the ability to relate to one who has not yet grokked, which has arguably devastating intellectual consequences when it comes time for them to develop a syllabus or write a textbook for those yet-to-have-grokked ones.
This is a bit tongue-in-cheek, so I want to make myself explicitly clear: backwards compatibility is not easy. The ability to put oneself in the shoes of someone who has not made some set of connections that one relies upon in their everyday life is a skill that must be honed.
However, many people, including many who teach on a regular basis, don’t bother to hone it, resulting in one of the many shared experiences between undergrads and nine-year-olds.
Mutual language in cross-cultural communication
Our ability to grok one another relies on our ability to establish and maintain a mutual language. Our ability to establish and maintain a mutual language relies on our ability to identify shared subgraphs and build outwards from them. This is easy when we share many subgraphs, and easier when those subgraphs are themselves expansive and densely connected.
The biggest challenge to cross-cultural communication, however, is not the relative challenge of establishing a mutual language when one does not share many densely connected subgraphs, but rather the prevalence of semantic incongruity.
There is no rule that states that a collection of nodes must correspond to the same set of edges and edge weights. It can certainly be more likely that some connections will form and not others depending on the terrain, but guarantee is at best an illusion. My direct path between A and B may route through C, D, and E to my conversational companion, but until these structures are made explicit, it is entirely possible that my companion and I will remain totally ignorant to the infrastructural differences between us.
Thus, when evaluating a system for a mutual language, we must examine not only the graphical similarity between two agents but also, and perhaps more importantly, the relationship between graphical similarity and graphical dissimilarity.
Cohesion and Flow
We communicate with our surroundings as we communicate with one another. If the paths are clear, we travel from place to place without interruption, freeing up some cognitive capacity in the process. Achieving “flow” or experiential cohesion is more often than not a result of an active absence (of barrier) rather than a concrete presence (of facilitator).
Rather than considering this to be the model of a “single-agent” system, I prefer to view it as a distributed and multimodal conversation, as this allows us to naturally extend the concept of the mutual language to nontraditional domains. The concept of a map is relevant here, and the concept of a generic infrastructure that can be laid over a real plot of land, assuming the region’s topography is compatible, and the concept of a perfect playlist for a given moment despite the fact that we cannot access these moments perfectly in advance. More on this later.
The gist of it
Nodes without structure float freely and clog the pipes.
When building computational models of “understanding,” it’s a good idea to grok our grokkage before teaching others to grok.
We should probably throw out the notion that anything’s meaning is inherent, or finite, or consistent.
Everything is what it does.
A future post will detail the benefits of pre-loaded mutual language vs fully encapsulated mutual language and the ways in which our choices between the two shape the entire information landscape.
A particularly poignant depiction of the importance of establishing a mutual language can be found in the infamous “Outliers plane crash chapter” — a section of Malcolm Gladwell’s book Outliers that explores the role of culture and communication in flight safety. In particular, the transcript from Avianca Flight 052 is worth a read.
For more context, google “Distributional Hypothesis” and read something that looks reputable. Though, it is worth noting that the analogy here is more of a cousin than a twin, and that the parallel occurs on a layer of abstraction beyond the particulars of distributional semantics.
See: “Google-Fu.”