Second Stanza: Of Phones and Phonemes

That the creative processes of all writers—poets, novelists, academics, and the like—are completely hidden, is what makes NLP so frustrating. Such underdetermination is the reason why XKCD can (justifiably) dish beeves upon computational linguists with such pizazz. There’s just no way to know what the hell is going on underneath the hood with human language, and all we have to go on is what comes out on the other side.

On that note, my Plan B for this project is a magical horse named “Zellig Horsis”. Give him any subject matter, anything at all, and that thoroughbred synthesized lyrical genius will tap a Pulitzer-winning poem in morse code for you. Put peanut butter under his lips, and you can even imagine he’s actually saying it!

Horses aside, with all of the million ways I could approach computational poetry, my inner-linguist compels me to seriously consider my formal training. My roots. “Consult the Kazan School, my son” suggests my inner-voice, sounding suspiciously like Jeff Goldblum.

Two Things I Learned in Phonology

Feature Analysis of Consonants

One of the great linguistic traditions is the Prague Linguistics Circle from the 1930’s. Among their many contributions was Feature Analysis, an entirely different way to understand phonetic inventories. It did so by looking inside the phonemes themselves; breaking through the seeming atomicity of phonemes to understand them as bundles of descriptive features.

For instance, consider the phoneme [d], as in :

Although the familiar articulatory perspective is more concise, feature analysis has the ability to take any two sounds and describe how they differ in much finer granularity. By breaking down phonetic strings into these features, I could construct a much more complex machine learning strategy. For instance, it could computationally determine the efficacy of sounds coming together inside words to compute a sort of poetic “score”.

Much of the delight in poetry comes from the delicious way sounds crash and coalesce to form a unique identity. Poetry, after all, is to be enjoyed out loud; the layering of semantic and prosodic and phonetic imagery is the essence and ultimately the mystique of poetry.

Articulation of Vowels

Vowels, as you might expect, are quite a bit different from consonants. Whereas consonants have a more distinct point of articulation (whether bilabial or dental or velar or uvular, placement variation doesn’t matter too much within that location), vowels are quite a bit harder to pin down. With continuous airflow through your vocal tract, you become a musical instrument of sorts. That is, the shape of your mouth changes what frequencies are generated and how it sounds.

If you ever took an undergraduate class in Linguistics, you’ve probably seen the infamous “tongue video”. It’s rough, but it does a good job of showing just how the tongue affects the shape of the oral tract, and how that corresponds to each phoneme:

Thing is, between languages like Hungarian—which has 14 distinct vowels—and the combination of neighboring consonants, which color a vowel, there are a lot of possible sounds for each “vowel”. That’s when I got this completely random insight: what if I used the same geometric algorithm to calculate vowel proximity that I would use to calculate cartographic entities? Since I’m already using MySQL, why not add a spatial column to map the geometry of each word? It might be just crazy enough to work.

Grasping at Straws

Where either of these approaches gets me remains to be seen. Like most things with NLP, it’s a crap shoot until you actually try it.

In the meantime, I’ve had fun writing Ruby glue code to pretty-print IPA. My next step is to implement a formal subclass or maybe independent analog of the String class that stores IPA strings as an array of phonemes, which each with their own feature bundle definition. Yes, it’s nerdy, but damn is Ruby meta-programming fun.