wagoner 1.0

A random word generator.

wagoner is an anagram for rawogen, a random words generator.

wagoner generates random words based on a given text. A random word is generated incrementally, and the given text is used to know which character should follow.

Usage

wagoner generates random words from tables extracted from texts. You need first to extract such a table from a text. To extract a table, use wagoner.table:

python -m wagoner.table TEXT --output text.table

This command will extract the table corresponding to TEXT into the file text.table.

Then, given the extracted table, wagoner.word can extract random words:

python -m wagoner.word text.table

This command will generate ten words of ten characters, based on the information in text.table. To generate another number of words, use the --count option and to change their length, the --length option.

Random words can also be generated from trees. While the tables only tell the generation which character should follow, trees also ensure that the generated word will start like a regular word (from the text it comes from) and will end like a regular (maybe different word). Trees are built from a table or from a text:

python -m wagoner.tree CONTENT --output text.tree

This command will extract a tree from CONTENT and save it into the file text.tree.

Warning: trees can be very large and expensive to build; to control their complexity, you can use the --prefix and --length options. See below for more information.

How does it work

The main idea behind wagoner is to use a text to conduct the generation of random words. The words in the text show what usually follows a given prefix, and this information is used to choose the next character when building a word incrementally.

For example, given the text composed of the two words ambiguous and gamblings, and the prefix ig, the first word tells us that the next character can be u (because in ambiguous, the sub-word ig is followed by u), and the second word tells us that the next character can be a or s (because in gamblings, the prefix g is followed by a or s). The next character is chosen randomly among the possible next characters and the word is incrementally generated following the same rule.

This next character is chosen randomly among the possible next characters, but weights are used to more probably use frequent characters, and longer prefixes. In other words, when choosing the next character for the prefix ig, more weight is given to u than to a and s because u follows ig in ambiguous while a and s only follow g in gamblings. Furthermore, if in the given text, a character follows more frequently a given prefix than another, the first one will have a higher chance to be chosen. This last characteristics of the generation can be turned down by using the --flatten option of wagoner.word. If this option is active, two characters following the same prefix will have the same chance to be chosen; nevertheless, a character following a shorter prefix will have less chance to be chosen than a character following a longer prefix.

This information of the frequence of successing characters in the words of the given text is stored in tables. Note that the --flatten option is also present with wagoner.table; this forces wagoner.table to extract a table in which the number of occurrences of a character following a given prefix is forgotten, giving the same result as the --flatten option of wagoner.word at generation.

Finally, when incrementally generating a random word, the full prefix is used to extract next character. In usual languages such as English, knowing the full prefix is not necessary to generate a pronounceable word; instead, three or four characters are usually sufficient to choose a next character that will sound correctly. Restricting the generation of words based on a bounded prefix instead of the full word can be achieved by using the --prefix option of wagoner.word. Similarly, the same option can be used on wagoner.table to directly generate (smaller) tables that will not propose next characters for longer prefixes.

Trees store more information than tables, but are also tailored to more strict generation. Trees store the information about which character should follow a given prefix, but only for words of a given length. They also store information to generate words starting as words of the text and ending as them. Because of this addition information, trees can be really complex, huge and long to generate; they can potentially store each word of a given length generable from the table. To tackle this complexity, trees can be limited to looking at only a suffix of the current prefix by using the --prefix option; furthermore, by building trees for shorter words (using the --length option), these trees are smaller.

Example

Let’s take the list of English words available on http://www-01.sil.org/linguistics/wordlists/english/. This list contains more than 100000 English words and should be useful to generate random words that sound English.

First thing to do before generating random words is to extract the table:

python -m wagoner.table wordsEn.txt --output wordsEn.table

The table is extracted and saved in the wordsEn.table file. Then, we can generate random words:

python -m wagoner.word wordsEn.table

that outputs, for example:

joggeriest
lvinistica
cleiadicat
zenerousne
ulencering
mmanencien
zeratenesi
keynoteric
encientnes
crappinesa

You can see that some words are really like real words: crappinesa really looks like crappiness, a word of wordsEn.txt. To avoid generating such too real words, we can ask wagoner to only consider finite prefixes; this will avoid to be trapped in the increasing proability of ressembling the word crappiness as the word is generated:

python -m wagoner.word wordsEn.table --prefix=2

that outputs, for example:

keyelittat
retimcenve
quedectrot
fodcalitur
xcedission
queffliqui
eshedlerad
ficklapett
quatersous
sulationur

In this case, the words are still pronounceable, but are not words of the wordsEn.txt file.

If you want to generate words that start and end the same way words of the text start and end, you can generate a tree from the table:

python -m wagoner.tree wordsEn.table --output=wordsEn.tree --prefix=3

The tree can then be used to generate words:

python -m wagoner.word wordsEn.tree

that outputs, for example:

sinationso
disjudging
titualimal
avespolybd
prophology
japackersc
nonneappra
overedefra
oxideodore
wordshedro

Library usage

wagoner can also be used as a library. wagoner.table.Table represent tables and a table can be extracted from a text file with:

table = Table.from_words(wagoner.utils.extract_words(text_file))

Table.from_words accept two optional arguments:

• prefix: if greater than 0, the length of the prefixes to take into account when generating the word (default is 0).

• flatten: if True, the table is flattened and two successing characters for the same prefix will have the same weight (default is False).

From such a table, a random word can be extracted:

word = table.random_word(word_length)

where word_length is the length of the desired word. random_word accepts several optional arguments:

• prefix and flatten, as above.

• start: if True, the generated word starts as a word of the text the table is extracted from (default is False).

• end: if True, the generated word ends as a word of the text the table is extracted from (default is False). Warning: this option should not be used because it is very time consuming.

Furthermore, wagoner.tree.Tree represent trees and a tree can be extracted from a table with:

tree = Tree.from_table(table, word_length)

where table is a table like the one above and word_length is the length of the words the tree will produce. Like Table.from_words, Tree.from_table supports two optional arguments: prefix and flatten. In this case, the prefix argument is very important because, if set to 0 (the default value), the tree can be really huge, even impossible to stay in memory, and will need a lot of time to be built.

From such a te, a random word can be extracted:

word = tree.random_word()

Unlike the tables, trees contain the complete information to extract words (their length, whether flattening the weights) and random_word takes no optional argument (technically, it accepts any arguments, but will ignore them).