Generating naming languages
You may also be interested in these notes which describe the map generator part of the bot.
I wanted to produce something which was a step above the usual alphabetic soup of generated placenames, and which was capable of producing recognisably distinct languages. The initial idea was that different regions of each map would have different languages, but I abandoned this because it was too hard to make it clear that this was what was going on, while still having the languages themselves be interesting.
The problem is to generate something like what the constructed languages (conlang) community call a 'naming language'. This is a light sketch of a language, focusing purely on the parts which are necessary to produce names. So there's little to no grammar, but a good sense of what the language sounds like, and how it's written. An excellent resource on how to do this 'by hand' is Mark Rosenfelder's Language Construction Kit, which goes on to explain how to flesh out your language into something that can manage actual sentences.
I'm going to talk through building up the generator I produced, stage by stage. This doesn't match at all the order in which the code was developed - in many cases I started with a much simpler version of a stage, then later tore it out and replaced it with what I describe below. There were also several dead ends which I tried and rejected, such as starting with a proto-language, and applying historical sound changes. Instead you're going to get the 'clean' version that makes it look like I knew what I was doing all along.
First steps: nonsense syllables
Let's start by generating some nonsense syllables, using a very basic set of sounds. For consonants, we'll use a fairly minimal set: /p t k m n s l/ (I'll use slashes // to indicate when I mean sounds, rather than letters). These are a nice simple set to work with, because they cover a variety of sounds, but they're all very common across languages, and they can all be written unambiguously in the Roman alphabet. For vowels we'll use the standard /a e i o u/. Each syllable will be a consonant, then a vowel, then another consonant (CVC).
These are fine, and they could all be perfectly good English words (if spelled a little differently). However, by picking the most common and boring sounds, we've ended up a bit lacking in personality. Real languages have distinctive sounds which make them more recognisable. So let's swap in some more varied consonant sets and see how things change.
As we add more sounds, we run into the problem of sounds which are difficult to spell in an unambiguous way, including a few sounds which don't exist in English. We'll deal with spelling properly later, so for now we'll use the International Phonetic Alphabet. For most sounds, the IPA symbols are the same as the usual Roman letters, but there are a few which are different:
- /ʃ/ - the 'sh' sound
- /ʒ/ - sometimes written 'zh', this is the 's' from 'pleasure', or a French 'j'
- /ʧ/ - the 'ch' sound, as in 'chair'. As the symbol suggests, this is actually just /t/ and /ʃ/ ('sh') mashed together
- /ʤ/- the 'j' from 'judge'. Again, it's really just /d/ and /ʒ/ ('zh') said at almost the same time
- /ŋ/ - the 'ng' sound from the end of 'hang'. This is a sound in its own right, and is really no more related to 'n' than 'ch' is to 'c'
- /j/ - the 'y' from 'year' (not to be confused with /ʤ/)
There are also a few sounds which might be unfamiliar. We don't want to go too far afield, or the reader will have no idea how to pronounce things, but adding a few extra sounds can go a long way to making something seem realistic.
- /x/ - this is a 'kh' sound, like in German 'Bach' or Scottish 'loch'
- /ɣ/ - a 'gh' sound, like /x/ but with the vocal chords vibrating. Can be heard in the Spanish 'amigo'
- /q/ - this is like a /k/, but pronounced further back in the mouth. It's probably best known from Arabic words like Qu'ran and Iraq
- /ʔ/ - the glottal stop. This does exist in English, but we never write it, so it can be hard to spot. It's the sound in the middle of 'uh-oh', or which replaces the 'tt' in 'button' in some accents.
As well as having a more varied selection, we can weight the consonants so that some are more common than others. One easy way of doing this is to order the consonants randomly, then when we have to choose one, we generate a random number between zero and one, square it (so it's still between zero and one, but biased towards the low end) and use that to index into the list. We'll re-use this technique later on whenever we have to randomly choose from a list - it gives the random choices a little bit of extra texture.
This is starting to make some progress. The Hawaiian sounds definitely have a different feel to the Arabic ones. However, these syllables all look a bit uniform. That's not necessarily bad within a language, but it does limit the amount of variation we can have between languages.
What we need is a better way of constructing syllables. This is the area of linguistics called phonotactics. Different languages combine their sounds in different ways, and this turns out to be responsible for a lot of the way a language feels. For example, let's look at a list of cities in China.
Even if you've never heard of them before, these feel like a unified set of names. The secret isn't in the sounds, which are fairly diverse and not very distinctive (at least to read). Instead it's in how they're put together.
First off, the majority of words in Mandarin have two syllables. This gives it a very clear rhythmic quality. Secondly, the syllable structure is pretty much constant. Each syllable consists of a single consonant, a vowel or two, and then optionally either /n/ or /ŋ/ (written ng). So 'shan' is a valid Mandarin syllable, but 'stan' and 'nash' are not.
English phonotactics are much more complicated, with up to three consonants at the start of a syllable ('street', 'spring'), and up to four at the end ('worlds', 'prompts'). There are also lots of restrictions on what can go where (e.g. no word starts with /ŋ/ or ends in /h/ - 'hang' is a word, but 'ngah' isn't). Replicating a system of that kind of complexity probably isn't worth the effort for our purposes, so all of the languages we make will be phonotactically fairly simple.
We'll introduce some new groups of consonants, sibilants (/s ʃ f/, etc), liquids/glides (/l r w j/) and finals (anything goes here, they're just a special 'end-of-syllable' group). Where before we had CVC syllables (consonant-vowel-consonant), we can now have things like S?CVF (optional sibilant-consonant-vowel-final).
So, adding some phonotactics gives us both variety (between languages) and consistency (within a language), both of which are excellent features. Unfortunately, as we've got it set up, there are a few problems. Sometimes we come up with syllables like 'ssot', which has the same sound twice in a row. Generally speaking this isn't something which happens in natural languages - the two sounds will fuse into a longer version of the same sound. We'd like to rule that out, along with some other cases of hard to pronounce clusters. A little bit of difficulty is fine, but too much and it just looks like we're throwing letters together at random.
The easiest thing to do here is just to check for these 'restricted' sequences of sounds, and if we find one, throw out the syllable and generate again.
At this point we're probably a bit sick of looking at arcane IPA symbols, and we're starting to think about how our language will be spelled. For the most part, it's not worth doing anything too fancy here. Real languages have complicated spelling systems with silent letters and context-dependent choices, but a lot of that will be completely invisible to a reader who has no idea how things should be pronounced in the first place. Instead, we want to focus on the aesthetics of the spelling system (or 'orthography' in linguistics terminology).
For example, looking at China again, the city of 广州 has been variously romanized as Kwang-chow, Kuang-chou, Kuaon-tseu, Kong-ziu, Gwong-zau, Gwóngjāu, Goangjoou or Guǎngzhōu. These all represent more or less the same underlying sounds, but give very different aesthetics.
Many sounds, of course, have very clear choices of spelling. Almost every language with an /m/ sound romanizes it as 'm', and vice versa, if you see an 'm' in an unknown language, you can be pretty sure it's pronounced as /m/. We'll start with a 'default' orthography, with standard correspondences like this, and apply 'patches' to it which cover the more ambiguous sounds. It would be easy to end up with a mishmash of conflicting symbols, but we'll get a much stronger sense of a unified aesthetic if we apply these patches in sets. For example, a French-inspired orthography would use 'ch' for /ʃ/ and 'j' for /ʒ/, while a Slavic-inspired set might use 'š' and 'ž'.
With consonants out of the way now, we can move on to vowels. We're going to cheat a little bit here. Whereas consonants are discrete and recognisable, vowels are kind of mushy and ambiguous. The sounds that are represented by 'a' in various languages are all sort of similar, but it's impossible to know exactly which one we're dealing with unless we know the language.
Because of this, we're not going to bother assigning phonetically accurate values to each vowel. Instead we'll have the basic /a e i o u/, and then some variants, which I'll write /A E I O U/ (these are not actual IPA symbols). Each language will have some subset of these, and we can apply a vowel orthography in the same way we did with the consonants.
Now that we've got some good syllables, we can start mushing them together to form words. This is another opportunity to give the language some distinctive flavour. We already saw how Mandarin's two-syllable words give it a clear aesthetic - there are also languages with extremely long words, like Greenlandic toqutserusuppoq 'to consider murder' or German Niederschlagswahrscheinlichkeit 'likelihood of rain'. Our language should have a typical range of word-lengths, bearing in mind that most languages have plenty of short words.
Nice, we now have a bunch of nonsense words with a consistent feel. If we wanted, we could happily stop here. However, there are a few extra things we can do to give our language a little bit of higher-level structure, and to help build the illusion of meaning.
In real languages, words are built out of units called 'morphemes', which are the smallest units of language which still carry meaning. Some words are single morphemes (e.g. 'happy', 'sad', 'platypus'), and some can be divided up into smaller units (e.g. 'un-help-ful', 'rest-less', 're-consider-s'). The relationship between syllables and morphemes, both of which combine to form words, is hazy at best.
Fortunately, we're making up our language, so we can make it as simple as we like. One syllable equals one morpheme. This lets us keep all our existing infrastructure. When we want to generate a word from a particular class, e.g. city names, we'll use one 'city' morpheme, and draw the rest from a pool of generic morphemes.
How do these pools of morphemes work? We want them to start empty, and to grow in a controllable way as we draw more and more samples. To do this, we keep a list of already created morphemes, and when choosing a morpheme to use, we select randomly from this list, plus a fixed number of 'new morpheme' tokens. If one of these tokens is chosen, then we generate a new syllable and add it to the list.
For generic morphemes, we use ten 'new morpheme' tokens, whereas for the word-class pools, we use just one. This means that, for example, once three morphemes have been generated, the word-class pools have a 3/4 chance of reusing an existing morpheme, whereas the generic pool only has a 3/13 chance. This gives a lot of variety in the generic morphemes (but still some reincorporation), while the class-specific morphemes will recur a lot. We also keep track of all morphemes so that we don't accidentally generate the same morpheme in two different pools (or twice in the same pool!).
We can use these words unmodified as names, but there are a few extra tricks which can add some texture to the language. Firstly we want to look at multi-word names. As a simple approach, we just decide that half the time, we're going to generate two words and combine them somehow. We can also have an optional one-syllable word which links the two, indicating some sort of preposition or grammatical particle. Multi-word names can be either joined by spaces or hyphens, depending on the orthography. We also allow a small proportion of names to start with a similar one-syllable grammatical particle (perhaps meaning 'the' or something similar).
Another trick we can apply is to re-use the morpheme pool idea at the word level. This means that multi-word placenames will often reincorporate the names of other places, which is nice. We also include a simple filter which prevents duplicate names.
Finally, we filter names based on length. Names which are too short are dull and repetitive, and names which are too long don't look great on the map. We simply reject anything with less than 5 characters, or more than 12.
So that's the version as currently running (more or less). There's code you can play with on GitHub. Here are some ideas which I never properly implemented, but which could be fun as extensions:
- Synchronic sound changes: rules which change some of the sounds depending on context (e.g. German /d/ becomes /t/ at the end of a word, Latin /k/ becomes /ʧ/ before /i/ or /e/)
- Historical sound changes: start with an 'ancestral form' and evolve the word systematically. Do this in multiple ways to produce a family of languages.
- Generated scripts: why should a generated language use the Roman alphabet? How do these people write things among themselves? Invent an alphabet (or similar) and show how the names are really spelled.
- Better grammar: break our morphemes out into roots and modifiers, and come up with some rules about how they're combined
If you do any of these things, or produce anything else interesting with the code, please let me know!
If you've enjoyed this piece, please consider contributing on Patreon so I can do more things like this.