Fictional languages are an aspect present in many hand-authored narratives, but underexplored in procedurally generated narratives. This paper presents an initial exploration of procedurally generated constructed languages (conlangs), following principles of language construction that relate phonology, morphology, and syntax. We present this generator as a tool for game developers. Our approach allows procedurally generated worlds and narratives to draw upon a generative space of languages that follow plausible linguistic principles.
spoken words. Those who understand more than one language will know that sentences or words may be strucResearch and game development has come a long way tured diferently, or that certain information is required for procedural generation of detailed game worlds grammatically in one language but entirely optional in that support rich narrative experiences, such as Dwarf others. These factors all contribute to the aesthetic of Fortress [1], with its interconnected systems of geology, the language, and can make a language feel out of place ecology, and history that inform how civilizations de- in certain contexts or appropriate to others. We would velop and interact [2]. These worlds often include simula- like for game developers that use this tool to be able to tions of human-like societies that develop social identity, supply parameters so that it creates a language with a culture, and tradition. However, the spoken and written iftting aesthetic for the context it will be used in. languages of these cultures remain an under-explored fea- Another goal for this project is to mimic patterns in ture for procedural world generation. In Dwarf Fortress, natural human languages so that it can generate lanfor example, the set of four languages (one for each race) guages that players perceive as believable. Therefore, the is completely fixed and only varies in terms of alphabet creation of the generator is informed by research in the and lexemes (i.e., each of the languages is a relexifica- ifeld of linguistics and is thus separated into sections that tion [3] of the others). deal with phonology, morphology, and syntax. Where
We propose that constructed languages, or conlangs possible, the models in the generator were informed by for short, ofer a window into how plausible languages empirical data. This mimicry has the added benefit of might be generated for fictional societies in games. In this making the tool easier to understand in terms of intupaper, we describe a work-in-progress generator based itions about human languages, as well as allowing the on documented conlang methods that can create a new generator to make decisions based on data from real huand unique conlang each time generation is invoked. We man languages. present the generator in its current form as a tool for The vision for this tool is so that it can be used in games game developers. in such a way that players will encounter a diferent con
One specific consideration for language generation in lang each time they play. One potential application is the context of game-worlds is the aesthetics of the lan- language-decoding puzzle gameplay (such as that found guage. For example, people who have a lot of exposure in Heaven’s Vault [4]) based on the conlang, in a way to languages with shared aesthetic features will be able that is robust to replay and to having answers posted onto pick up the sound patterns and tendencies of those line. However, there are also more general opportunities languages and be able to distinguish that language from to have the language of a generated civilization reflect a lineup without necessarily actually understanding the aspects of its culture, to include the writing in in-world artifacts (such as signage and books), and to see it written or spoken by NPCs. In the long term, we imagine this tool supporting games that include procedurally generated worlds and narratives enriched by unique languages.
AIIDE Workshop on Experimental Artificial Intelligence in Games, October 08, 2023, University of Utah, Utah, USA $ cai.a@northeastern.edu (A. Cai); c.martens@northeastern.edu (C. Martens)
© 2023 Copyright for this paper by its authors. Use permitted under Creative Commons License CPWrEooUrckReshdoinpgs IhStpN:/c1e6u1r3-w-0s.o7r3g ACttEribUutRion W4.0oInrtekrnsahtioonpal (PCCroBYce4.0e).dings (CEUR-WS.org) In the course of this paper, there will be discussion about potential game developers who could use this tool to 4. Phonology assist in the creation of games, as well as players who could play those created games. For the purpose of clar- The phonology of a language is all the sounds, or ity, game developers will be henceforth referred to as phonemes, that speakers of the language use when speakdevelopers, and game players will be referred to as play- ing that language. Phonology can difer greatly between ers. It is important to note that developers will not refer languages and is one of the first things to be noticed to us, the authors and creators of the generator. We will by someone who does not understand a language beavoid the term “users”, even though it may be natural to ing spoken. Therefore, take phonology as the biggest contributing factor to the aesthetic of a game-world language when the player first encounters it. In settings with verbalized NPC dialog, it would likely be the most salient aspect of the language (prior to the player learning the language’s structure and semantics). One aspect of phonology that is easily modeled digitally is the size of the sound inventory. This describes the amount of diferent sounds speakers of the language make while speaking that language. Languages can difer in this way by incredible magnitudes. Some languages have as few as 6 consonants, for example, while others have more than 100 [7]. As for which sounds are available for selection, most linguists use the International Phonetic Alphabet (IPA) [8] (see Figure 2). All sounds that have been observed by linguists to occur in any human language can be notated using IPA [9]. Looking through all the IPA symbols gives us a list of sounds to select from. It is of note that the IPA does not have a symbol for every possible sound a human mouth can make: there are a few sounds that are technically possible but have not been observed in any human language.
The overall structure of this macro-module is a linear path from one micro-module to the next (see Figure 3).
We start of with sound inventory size, then move onto sound selection. The output is a phonology object which stores all the sounds within the generated language.
If the developer does not specify a size for the phoneme inventory, the tool will choose one based on a distribution derived from a database [7] which surveyed real languages all around the world (see Figure 4). The distribution appears to be a normal distribution at first glance, but the categories are not equally divided. The “large” category actually includes a much larger range of numbers than the other categories. One could attempt to regress a skewed probability distribution, but picking a category with a distribution matching the data and then picking a number in that category in an uniform distribution approximates it suficiently.
Phonemes are modeled in the generator using objects that store information such as a long, descriptive name for ease of handling within the code, the IPA notation, and information about how the sound is created using the mouth. The latter information is not immediately important but will become relevant in other parts of the when the player is intended to learn to speak or write the language.
In our implementation, the morphology macro-module is less linear than the phonology module. It consists of three micro-modules: morpheme, syllable, and phonotactics. The primary function lies in the morpheme micromodule, which uses the syllable and phonotactics micromodules in its execution. The output is a morphology object which contains the outputs of both the phonotactics and morpheme micro-modules (see Figure 5). generator. These objects can be of either the consonant type or the vowel type, but both will eventually make up a phoneme object. The reason why this additional level of abstraction is necessary is due to the fact that some languages treat multiple consonants or vowels as a singular phoneme. For example, the /tS/ sound in English is treated as a single phoneme by English despite typically being spelled with two alphabetic consonants (“ch”) and being composed of two IPA sounds. The generator will also exclude phonemes that have never been observed in any human language. This set is the same every time the generator is run, so the information is stored in a JSON document that the script reads upon execution.
In terms of actually selecting the sounds, some sounds
are more common than others. When the selection of
sounds is not specified by the developer, the generator
picks sound in a distribution informed by a database of
phonological data of real human languages [
Once all the phonemes are selected are selected, they are stored in a phonology object, which is in essence an array of phoneme objects with some wrapper functions.
The phonology object is returned and can then be used as the input for the morphology.
Syllables may seem like a concept that is the same across all languages, but diferent languages define syllables in diferent ways. Certain languages may define a series of vowels as a diphthong, which would count for only one syllable, while others count that same series of vowels as separate syllables (e.g., /aI/ as in “my” in English compared to /aI/ as in “愛” in Japanese).
In addition, consonants at the beginning of a syllable may have diferent rules than consonants at the end of a syllable. Diferent languages can have rules on how vowels appear in syllables as well. English can have three vowels as a triphthong in a single syllable as in the word “flour,” but other languages such as Japanese allows only a singular vowel sound in a syllable. In addition, some languages, such as Mandarin, use tones. Tonal languages diferentiate syllables with the pitch or change in pitch of articulation.
Even with the overall structure of the syllable decided, 5. Morphology there may be additional rules on which combinations of sounds are allowed together. For example, even though Morphology is the study of the construction of words. English allows for a consonant cluster of two at the beWords are not easy to classify as they can difer in defi- ginning of a syllable, an “s” sound followed by an “r” nition or purpose between diferent languages [ 12]. For sound is not allowed. In addition, in many languages example, large numbers in English are expressed in what (including English), consonants can come in clusters that is defined as multiple words, but those same numbers in still contribute to only a single syllable, but the maxiGerman are expressed in what is defined as one word. For mum number of consonants in a single cluster difers by this reason, the tool does not deal with words but with language. syllables and morphemes, which are more concretely deifned in linguistics. Morphology is an aspect of language that becomes more relevant to game-world languages
For the sake of simplicity, the generator regards syllables as a sequence of vowels with potentially a sequence of consonants on either side. Due to the way that languages difer greatly in syllable construction rules, the generator encourages developers to specify how many consonants can be at the beginning and end of each syllable, as well as how many vowels can appear in the middle. They can also specify whether tones are diferentiating within the language, and how many and which tones are used.
Within a single language, certain syllable structures may be more common than others. For example, English speech has many more syllables starting with one or two consonants than with three. Due to lack of data on this matter, the generator uses a Zipfian distribution, which is a distribution observed in many natural distributions, such as frequency of a word within bodies of text, or population of cities. In a Zipfian distribution, the frequency of occurrence of an element is inversely proportional to its rank in a frequency table. The generator stores these syllables as an object which primarily consists of a sequence of phonemes. However, unlike phonemes, not all possible syllables within the language are stored as the number of possible syllables may be orders of magnitude higher than that of possible phonemes, depending on the rules of syllable construction.
Phonotactics refers to such rules as which sounds are allowed together, as well as how sounds can change in certain contexts [13]. For example, the vowel in the second syllable of “modal” changes when an additional part is added onto the end as in “modality.” There is an incredibly large array of such rules for change across real human languages [14]. We did not find a database documenting the frequency of such rules across languages, so we were not able to weigh the probability of phonotactic rules being picked by the generator by their frequency in real human languages. 5.2.1. Implementation rules, but that may be unnecessary for believability and is beyond the expertise of the authors. The phonotactics rules are stored as a phonotactics object consisting of multiple rules stored as strings. When a syllable is constructed, the syllable micro-module checks to make sure that every rule is followed.
There may be a more computationally eficient method of storing these rules than as strings, but they are easy to work with in development. An optimization pass in the future may change the way these are stored
Morphemes are defined by linguists as atomic units of meaning [15]. The “s” sound that English speakers attach onto the end of words to indicate plurality is an example of a morpheme. Evidently, morphemes can consist of multiple syllables, a singular syllable, or even a consonant or vowel that changes another morpheme’s syllable. Because they are so flexible, morphemes are stored as an object that consists of either phonemes or syllables, with a property indicating how they are to be fitted with other morphemes. This can specify that they are free, which means they can stand alone in speech and are easy to deal with. However, bound morphemes cannot stand alone and thus also store rules on how they are to be attached to other morphemes.
With the tools of syllable construction and phonotactic rules, we can begin to construct morphemes. In the morpheme micro-module, the generator can finally assign meaning to sounds. Morphemes are stored as objects with strings indicating its pronunciation and meaning, as well as an enum for its grammatical type. After a basic set of morphemes are created, they are stored in an morphology object, which also stores the phonotactics object from the previous submodule. The morphology object can then be passed on to the final output
The phonotactics micro-module includes a few common Syntax is the subdiscipline of linguistics that deals with phonotactic rules that we deemed worth emulating, e.g. the construction of sentences from morphemes [16]. Synthe tendency for unvoiced consonants to become voiced tax can difer between languages to a degree that may when they appear between multiple occurrences of the surprise those who speak one language. For example, same vowel. English adjectives typically appear before the noun it is
If the developer does not specify phonotactic rules, describing, but in Romance languages, adjectives typithe generator will pick several, but not all, at random cally occur after the described noun [17]. One easy aspect to be enabled. The generator will first pick a number of syntax to model digitally is word order, i.e. the order between a quarter to a half of the total number of rules things appear in sentences. English sentences typically available to pick from. It will then pick that many using a start with a subject, followed by the verb, and then the uniform distribution. There may be correlation between object. This order can be abbreviated to SVO (Subject, the nature of the sound inventory and the phonotactic Verb, Object). However, this is not the only possibility.
We are in the progress of implementing the design described in this paper as the GenLang tool available on GitHub at https://github.com/AkaiGameDev/GenLang. Currently, we have implemented the Phonology module and the Morphology module except for Phonotactics. Phonotactics and Syntax remain to me implemented.
Figure 8 shows the current command-line interface and corresponding output. Each prompt ending in “:” represents an opportunity for developer input. The system then prints out the generated phonology including a 2-dimensional consonant chart, sample morphemes generated and associated meanings. See Appendix A for a second input-output example in text form.
There are six ways to arrange an ordering of three distinct objects. Therefore, there are six possible ways to order the subject, verb, and object in a sentence. While all permutations have been observed in real human languages, some orders are far more common than others. Notably, it seems that vast majority of human languages have the subject occur before the object [18] (see Figure 6).
The syntax macro-module creates the rules for the construction of sentences (see Figure 7). While it may seem necessary for the macro-module to use the morphology macro-module, that is not actually the case. The syntax module is solely responsible for the rules with which morphemes or morphemes clusters are ordered in sentences, regardless of what the morphemes or morpheme clusters are comprised of. Syntax is something that can greatly increase the feeling of foreignness for players if they are learning to understand the language.
For sentence word order, we include both the common if a verb can be conjugated to encode the subject-pronoun (as in Spanish), the pronoun subject can be dropped (“pro drop”). Additionally, we do not treat aspects of vocabulary that might be afected by the cultural context of the world for which the language is being developed. Nor do we treat any aspect of semantics, e.g. a process for constructing character utterances appropriate to specific game-world contexts, although this task is one over which we assume a developer would usually want more manual control.
Finally, GenLang is a project by humans with inherent biases, so it cannot possibly avoid linguistic bias nor fully represent the diversity of human languages. Therefore, this generator comes with the disclaimer that the creators are not experts in linguistics, and (across all authors) are most familiar with English, Mandarin, Japanese, and Spanish (our “source languages”). There are going to be unaccounted-for aspects of real-world languages and linguistic topics that the authors are unfamiliar with. Fortunately, our source languages difer from each other in numerous ways and thus provide insight into how languages can difer. However, it is likely that the differences across our source languages are emphasized in the design of the generator and the similarities are overlooked. If others with an entirely diferent set of source languages and linguistic or computational knowledge were to tackle the challenge of algorithmically modeling languages, they may focus more on other aspects of languages, resulting in a diferent algorithm. In addition, the authors acknowledge that linguists’ understanding of languages is still changing, and that even models made by practicing linguists on well-studied real human languages are not without flaws. As this project replicates such models, it follows that the outputs of this generator may include such flaws as well.
Our main contribution is a preliminary algorithm, designed as a set of modules, for procedurally generating a constructed language for use in a game world. However, this project is very much a work in progress, and also entails some inherent limitations.
First, while we have done some preliminary exploration of written language generation, it is not incorporated into this generator and brings its own set of challenges. A written language generator would have to generate 2D artifacts with constraints that are dificult to proceduralize. Written languages typically have glyphs that are easy to write and distinguish from each other. 9. Conclusion It may be dificult to algorithmically evaluate how well generated artifacts satisfy those constraints. A written We have presented a proof-of-concept generator for language generation could be integrated with our genera- phonology, morphology, and syntax in a constructed tor to allow for a more sophisticated generated language, language, drawing from empirical linguistics research as well as supporting gameplay with more emphasis on to represent a cross-section of the possibility space of visual communication than audio (e.g. text games). languages spoken on Earth. Taking into account the in
Second, there are important aspects of syntax and se- tended use case of generating languages within fictional mantics that we have not yet implemented, or in some game worlds, we have presented our design decisions cases designed good solutions for. For example, the for this tool. We demonstrate this design with a working problem of generating a lexicon (set of meanings to as- software implementation.1 As a result, this project makes sign to morphemes) is currently the responsibility of the significant inroads towards the longer-term ambition of developer, and currently lexicon meanings will be ran- generating complete, novel constructed languages that domly assigned to morphemes. But this approach does incorporate developer input. not take into account the possible relationships between morphemes, including syntactic concerns like stemming, inflection, conjugation, tense, aspect, and mood. These choices could impact other aspects of syntax; for example, 1Available at https://github.com/AkaiGameDev/GenLang.
IPA character: m Descriptive name: VoicedBilabialNasal Place of Artic.: [’bilabial’]
[’nasal’] IPA character: s Descriptive name: VoicelessAlveolarSibilant Place of Artic.: [’alveolar’] Manner: [’sibilant’] IPA character: b Descriptive name: VoicedBilabialPlosive Place of Artic.: [’bilabial’] Manner: [’plosive’] IPA character: n Descriptive name: VoicedAlveolarNasal Place of Artic.: [’alveolar’] Manner: [’nasal’] IPA character: k Descriptive name: VoicelessVelarPlosive Place of Artic.: [’velar’] Manner: [’plosive’] IPA character: t Descriptive name: VoicelessAlveolarPlosive Place of Artic.: [’alveolar’] Manner: [’plosive’] IPA character: j Descriptive name: VoicedPalatalApproximant Place of Artic.: [’palatal’] Manner: [’approximant’] IPA character: d Descriptive name: VoicedAlveolarPlosive Place of Artic.: [’alveolar’] Manner: [’plosive’] IPA character: g Descriptive name: VoicedVelarPlosive Place of Artic.: [’velar’] Manner: [’plosive’] IPA character: f Descriptive name:
VoicelessLabiodentalFricative Place of Artic.: [’labiodental’] Manner: [’fricative’] IPA character: ä Descriptive name: VoicedVelarImplosive Place of Artic.: [’velar’] Manner: [’implosive’] IPA character: õ Descriptive name: VoicedRetroflexApproximant Place of Artic.: [’retroflex’] Manner: [’approximant’] IPA character: ú’ Descriptive name: RetroflexEjectiveStop Place of Artic.: [’retroflex’] Manner: [’plosive’, ’ejective’] IPA character: Q Descriptive name: VoicedPharyngealFricative Place of Artic.: [’pharyngeal’] Manner: [’fricative’] IPA character: ù Descriptive name: VoicelessRetroflexSibilant Place of Artic.: [’retroflex’] Manner: [’sibilant’] IPA character: &’ Descriptive name: PalatalEjectiveFricative Place of Artic.: [’palatal’] Manner: [’fricative’, ’ejective’] IPA character: q Descriptive name: VoicelessUvularPlosive Place of Artic.: [’uvular’] Manner: [’plosive’] IPA character: ü Descriptive name: VoicedRetroflexSibilant Place of Artic.: [’retroflex’] Manner: [’sibilant’] IPA character: p Descriptive name: VoicelessBilabialPlosive Place of Artic.: [’bilabial’] Manner: [’plosive’] IPA character: r Descriptive name“: VoicelessAlveolarTrill Place of Artic.: [’alveolar’] Manner: [’trill’] IPA character: g! Descriptive name: VoicedAlveolarClick Place of Artic.: [’alveolar’, ’velar’] Manner: [’click’] Vowels: i u e o Generated 3 sentences: English translation: The man bakes good bread Generated language: äa togi äosa kitijo English translation: The child eats bad bread Generated language: ju &’o üogibaso äosa English translation: The woman watches the child cook Generated language: ge relime ju togi “