I remade and improved my random name generator algorithm I had done in Ruby several years ago, but this time in F#.
It works by taking a sample file which contains names, the names should be thematically similar, and uses it to create chains of probabilities. That is, when we find the letter A in the sample, what are the possible letters that can follow this A and what probability is there for each of these letter to come up.
This probability chain can have any length bigger than one.
Here are the steps for this algorithm:
- Build a probability table from the input file.
- Generate name length info from the input file.
- Generate a name with the name length and probability table.
Build a probability table
Here’s what a probability table looks like:
{“probabilities”:{” “:{“al”:0.973913,”am”:0.965217,”ar”:0.93913,”at”:0.930435,”au”:0.921739,”ba”:0.904348,”be”:0.886957,”bi”:0.878261,”bo”:0.86087,”bu”:0.852174,”ca”:0.843478,”co”:0.834783,”da”:0.808696,”de”:0.791304,”do”:0.782609,”dr”:0.773913,”el”:0.756522,”eo”:0.747826,”fa”:0.73913,”ga”:0.721739,”gh”:0.713043,”gi”:0.704348,”gr”:0.695652,”gu”:0.669565,”ha”:0.643478,”ho”:0.626087,”is”:0.617391,”je”:0.6,”ju”:0.591304,”ka”:0.582609,”ko”:0.556522,”ku”:0.547826,”la”:0.53913,”li”:0.521739,”lo”:0.513043,”lu”:0.495652,”ma”:0.486957,”me”:0.478261,”mh”:0.46087,”mi”:0.452174,”mo”:0.426087,”no”:0.4,”on”:0.391304,”or”:0.373913,”pa”:0.356522,”ph”:0.330435,”pu”:0.321739,”qa”:0.313043,”qu”:0.304348,”ra”:0.278261,”rh”:0.269565,”ri”:0.26087,”ro”:0.234783,”ru”:0.226087,”sa”:0.191304,”se”:0.182609,”sh”:0.173913,”ta”:0.147826,”th”:0.13913,”to”:0.130435,”tu”:0.121739,”ul”:0.113043,”va”:0.086957,”vo”:0.078261,”wa”:0.069565,”wi”:0.06087,”xa”:0.052174,”xe”:0.043478,”yu”:0.026087,”ze”:0.017391,”zi”:0.008696,”zu”:0.0},”a”:{“ba”:0.970874,”be”:0.951456,”de”:0.941748,”di”:0.932039,”ev”:0.92233,”go”:0.893204,”gu”:0.883495,”hd”:0.873786,”hr”:0.864078,”ie”:0.854369,”ig”:0.84466,”im”:0.834951,”
…
,”nameLengthInfo”:{“mean”:7.4273504273504276,”standardDeviation”:1.7261995176214626}}
This table can be serialized to prevent recomputing it each time we call the algorithm.
The algorithm works with sub strings of size X where a small X will provide more random results (less close to the original result) but a larger X will provide results more closely aligned with the sample file.
Results more closely aligned with the sample file better reflect the sample but face a higher risk of ending up as a pastiche of 2 existing names or in some cases being one of the sample’s name as is.
Here’s how the sub strings work:
If we have the following name in our sample file:
Gimli
Using a sub string length of 2 would add all these sub strings in our probability table:
G i
i m
m l
l i
While using a sub string length of 3 would add these sub strings:
G im
i ml
m li
And so on as we increase the length of the sub strings.
After we have counted all the possible occurrences of each sub string over the whole file we assign a probability to each one.
For example, if for our whole file we would have the following possible sub strings for G
G im
G lo
G an
Each of these would be assigned a probability of 33.3%.
Generate name length info from the input file
The generate names of a length representative of our input sample we simply count the length of each name and derive a mean value and standard deviation. Using the mean and standard deviation will then easily allow us to draw a value from the normal distribution of word lengths.
Additionally it’s best to enforce a minimum word length. Even if our sample contains shorter names (2 or 3 letters long), from experience the algorithm doesn’t produce convincing results on these shorter lengths.
This is because it doesn’t differentiate sub strings for long and short names.
Generate a name with the name length and probability table
To generate a name we start by finding our desired name length using our name length info. Then we select our first character, the white space character.
We then generate a number between 0.0 and 1.0 (or 0 and 100) and using a prebuilt dictionary containing the probability table, find the next item.
Code
The code is also available on GitHub in a more readable format.
Sample and Examples
The larger the sample the better. Also the more thematically aligned the sample, the better. What I mean by thematically aligned is if you include the names of all Greek masculine mythological figures, you will get results that resemble the names of the Greek heroes and Gods.
For example:
Thedes
Kratla
Pourseus
On the other hand if you build your samples with names from the Lord of The Rings but include an equal part of Hobbits, Dwarf, Elven and Orcish names you will end up with a mishmash that does not make much sense.
Finally here are some results of the algorithm using the this sample file containing the names of some of the locations in the games Final Fantasy XI and Final Fantasy XIV:
Bastok SanDoria Windurst Jeuno Aragoneu Derfland Elshimo Fauregandi Gustaberg Kolshushu Kuzotz LiTelor Lumoria Movalpolos Norvallen Qufim Ronfaure Sarutabaruta Tavnazian TuLia Valdeaunia Vollbow Zulkheim Arrapago Halvung Oraguille Jeuno Rulude Selbina Mhaura Kazham Norg Rabao Attohwa Garlaige Meriphataud Sauromugue Beadeaux Rolanberry Pashhow Yuhtunga Beaucedine Ranguemont Dangruf Korroloka Gustaberg Palborough Waughroon Zeruhn Bibiki Purgonorgo Buburimu Onzozo Shakhrami Mhaura Tahrongi Altepa Boyahda RoMaeve ZiTah AlTaieu Movalpolos Batallia Davoi Eldieme Jugner Phanauet Delkfutt Bostaunieux Ghelsba Horlais Ranperre Yughott Balga Giddeus Horutoto Toraimarai Lufaise Misareaux Phomiuna Riverne Xarcabard Gusgen Valkurm Ordelle LaTheine Konschtat Arrapago Carteneau Thanalan Coerthas Noscea Matoya MorDhona Gridania Rhotano Uldah Limsa Lominsa Dravanian Ishgard Doma Sastasha Tamtara Halatali Haukke Qarn Aurum Amdapor Pharos Xelphatol Daniffen Aldenard Garlea Eorzea Vanadiel
Note that this sample is very small and not thematically consistent, still here are the results using a sub string length of 2:
Gazormon
Vasamaur
Ltemenolp
Zonaone
Ldausa
Zorvaie
Kugo
Limoruxeab
Raullaiat
Jelphorshi
A sub string length of 3:
Arzergid
Mhaure
Phowindugh
Rhonearlos
Tuto
Saltabaolo
Qangarle
And a sub string length of 5:
Arronfais
Sautotara
Tahimorut
Batahranperr
Movernguegan
I feel that the algorithm could still use some improvements but is still very satisfactory considering the bad quality of the sample file used.