7.3 Phonetic Code

Aspell is in fact the spell checker that comes up with the best suggestions if it finds an unknown word. One reason is that it does not just compare the word with other words in the dictionary (like Ispell does) but also uses phonetic comparisons with other words.

The new table driven phonetic code is very flexible and setting up phonetic transformation rules for other languages is not difficult but there can be a number of stumbling blocks — that’s why I wrote this section.

The main phonetic code is free of any language specific code and should be powerful enough to allow setting up rules for any language. Anything which is language specific is kept in a plain text file and can easily be edited. So it’s even possible to write phonetic transformation rules if you don’t have any programming skills. All you need to know is how words of the language are written and how they are pronounced.

7.3.1 Syntax of the transformation array

In the translation array there are two strings on each line; the first one is the search string (or switch name) and the second one is the replacement string (or switch parameter). The line

version   version

is also required to appear somewhere in the translation array. The version string can be anything but it should be changed whenever a new version of the translation array is released. This is important because it will keep Aspell from using a compiled dictionary with the wrong set of rules. For example, if when coming up with suggestion for hallo, Aspell will use the new rules to come up with the soundslike say H*L*, but if ‘hello’ is stored in the dictionary using the old rules as HL instead of H*L* Aspell will never be able to come up with ‘hello’. So to solve this problem Aspell checks if the version strings match and aborts with an error if they don’t. Thus it is important to update it whenever a new version of the translation array is released. This is only a problem with the main word list as the personal word lists are now stored as simple word lists with a single header line (i.e. no soundslike data).

Each non switch line represents one replacement (transformation) rule. Words beginning with the same letter must be grouped together; the order inside this group does not depend on alphabetical issues but it gives priorities; the higher the rule the higher the priority. That’s why the first rule that matches is applied. In the following example:

GH   _
G    K

‘GH -> _’ has higher priority than ‘G -> K’

‘_’ represents the empty string “”. If ‘GH -> _’ came after ‘G -> K’, the second rule would never match because the algorithm would stop searching for more rules after the first match. The above rules transform any ‘GH’ to an empty string (delete them) and transforms any other ‘G’ to ‘K’.

At the end of the first string of a line (the search string) there may optionally stand a number of characters in brackets. One (only one!) of these characters must fit. It’s comparable with the ‘[ ]’ brackets in regular expressions. The rule ‘DG(EIY) -> J’ for example would match any ‘DGE’, ‘DGI’ and ‘DGY’ and replace them with ‘J’. This way you can reduce several rules to one.

Before the search string, one or more dashes ‘-’ may be placed. Those search strings will be matched totally but only the beginning of the string will be replaced. Furthermore, for these rules no follow-up rule will be searched (what this is will be explained later). The rule ‘TCH-- ’-> _ will match any word containing ‘TCH’ (like ‘match’) but will only replace the first character ‘T’ with an empty string. The number of dashes determines how many characters from the end will not be replaced. After the replacement, the search for transformation rules continues with the not replaced ‘CH’!

If a ‘<’ is appended to the search string, the search for replacement rules will continue with the replacement string and not with the next character of the word. The rule ‘PH< -> F’ for example would replace ‘PH’ with ‘F’ and then again start to search for a replacement rule for ‘F…’. If there would also be rules like ‘FO ’-> ‘O’ and ‘F -> _’ then words like ‘PHOXYZ’ would be transformed to ‘OXYZ’ and any occurrences of ‘PH’ that are not followed by an ‘O’ will be deleted like ‘PHIXYZ -> IXYZ’. The second replacement however is not applied if the priority of this rule is lower than the priority of the first rule.

Priorities are added to a rule by putting a number between 0 and 9 at the end of the search string, for example ‘ING6 -> N’. The higher the number the higher is the priority.

Priorities are especially important for the previously mentioned follow-up rules. Follow-up rules are searched beginning from the last string of the first search string. This is a bit complicated but I hope this example will make it clearer:

CHS      X
CH       G

HAU--1   H

SCH      SH

In this example ‘CHS’ in the word ‘FUCHS’ would be transformed to ‘X’. If we take the word ‘DURCHSCHNITT’ then things look a bit different. Here ‘CH’ belongs together and ‘SCH’ belongs together and both are spoken separately. The algorithm however first finds the string ‘CHS’ which may not be transformed like in the previous word ‘FUCHS’. At this point the algorithm can find a follow-up rule. It takes the last character of the first matching rule (‘CHS’) which is ‘S’ and looks for the next match, beginning from this character. What it finds is clear: It finds ‘SCH -> SH’, which has the same priority (no priority means standard priority, which is 5). If the priority is the same or higher the follow-up rule will be applied. Let’s take a look at the word ‘SCHAUKEL’. In this word ‘SCH’ belongs together and may not be taken apart. After the algorithm has found ‘SCH ’-> ‘SH’ it searches for a follow-up rule for ‘H+’‘AUKEL’. It finds ‘HAU--1 -> H’, but does not apply it because its priority is lower than the one of the first rule. You see that this is a very powerful feature but it also can easily lead to mistakes. If you really don’t need this feature you can turn it off by putting the line:

followup      0

at the beginning of the phonetic table file. As mentioned, for rules containing a ‘-’ no follow-up rules are searched but giving such rules a priority is not totally senseless because they can be follow-up rules and in that case the priority makes sense again. Follow-up rules of follow-up rules are not searched because this is in fact not needed very often.

The control character ‘^’ says that the search string only matches at the beginning of words so that the rule ‘RH -> R’ will only apply to words like ‘RHESUS’ but not ‘PERHAPS’. You can append another ‘^’ to the search string. In that case the algorithm treats the rest of the word totally separately from the first matched string at the beginning. This is useful for prefixes whose pronunciation does not depend on the rest of the word and vice versa like ‘OVER^^’ in English for example.

The same way as ‘^’ works does ‘$’ only apply to words that end with the search string. ‘GN$ -> N’ only matches on words like ‘SIGN’ but not ‘SIGNUM’. If you use ‘^’ and ‘$’ together, both of them must fit ‘ENOUGH^$ -> NF’ will only match the word ‘ENOUGH’ and nothing else.

Of course you can combine all of the mentioned control characters but they must occur in this order: ‘< - priority ^ $’. All characters must be written in CAPITAL letters.

If absolutely no rule can be found — might happen if you use strange characters for which you don’t have any replacement rule — the next character will simply be skipped and the search for replacement rules will continue with the rest of the word.

If you want double letters to be reduced to one you must set up a rule like ‘LL- -> L’. If double letters in the resulting phonetic word should be allowed, you must place the line:

collapse_result     0

at the beginning of your transformation table file; otherwise set the value to ‘1’. The English rules for example strip all vowels from words and so the word "GOGO" would be transformed to "K" and not to "KK" (as desired) if collapse_result is set to 1. That’s why the English rules have collapse_result set to 0.

By default, all accents are removed from a word before it is matched to the soundslike rules. If you do not want this then add the line

remove_accents      0

at the beginning of your file. The exact definition of an accent is language dependent and is controlled via the character set file. If you set remove_accents to ’0’ then you should also set "store-as" to "lower" in the language data file (not the phonetic transformation file) otherwise Aspell will have problems when both the accented and the de-accented version of a word appearing in the dictionary; it will consider one of them as incorrectly spelled.

7.3.2 How do I start finally?

Before you start to write an array of transformation rules, you should be aware that you have to do some work to make sure that things you do will result in correct transformation rules.

7.3.2.1 Things that come in handy

First of all, you need to have a large word list of the language you want to make phonetics for. It should contain about as many words as the dictionary of the spell checker. If you don’t have such a list, you will probably find an Ispell dictionary at http://fmg-www.cs.ucla.edu/geoff/ispell-dictionaries.html which will help you. You can then make affix expansion via ‘ispell -e’ and then pipe it through ‘tr " " "\n"’ to put one word on each line. After that you eventually have to convert special characters like ‘é’ from Ispell’s internal representation to latin1 encoding. ‘sed s/e'/é/g’ for example would replace all ‘e'’ with ‘é’.

The second is that you know how to use regular expressions and know how to use grep. You should for example know that:

grep ^[^aeiou]qu[io] wordlist | less

will show you all words that begin with any character but ‘a’, ‘e’, ‘i’, ‘o’ or ‘u’ and then continue with ‘qui’ or ‘quo’. This stuff is important for example to find out if a phonetic replacement rule you want to set up is valid for all words which match the expression you want to replace. Taking a look at the regex(7) man page is a good idea.

7.3.2.2 What the phonetic code should do

Normal text comparison works well as long as the typer misspells a word because he pressed one key he didn’t really want to press. In these cases, mostly one character differs from the original word.

In cases where the writer didn’t know about the correct spelling of the word, the word may have several characters that differ from the original word but usually the word would still sound like the original. Someone might think that ‘tough’ is spelled ‘taff’. No spell checker without phonetic code will come to the idea that this might be ‘tough’, but a spell checker who knows that ‘taff’ would be pronounced like ‘tough’ will make good suggestions to the user. Another example could be ‘funetik’ and ‘phonetic’.

From these examples you can see that the phonetic transformation should not be too fussy and too precise. If you implement a whole phonetic dictionary as you can find it in books this will not be very useful because then there could still be many characters differing from the misspelled and the desired word. What you should do if you implement the phonetic transformation table is to reduce the number of used letters to the only really necessary ones.

Characters that sound similar should be reduced to one. In the English language for example ‘Z’ sounds like ‘S’ and that’s why the transformation rule ‘Z -> S’ is present in the replacement table. “PH is spoken like “F and so we have a ‘PH -> F’ rule.

If you take a closer look you will even see that vowels sound very similar in the English language: ‘contradiction’, ‘cuntradiction’, ‘cantradiction’ or ‘centradiction’ in fact sound nearly the same, don’t they? Therefore the English phonetic replacement rules not only reduce all vowels to one but even remove them all (removing is done by just setting up no rule for those letters). The phonetic code of “contradiction” is “KNTRTKXN” and if you try to read this letter-monster loud you will hear that it still sound a bit like ‘contradiction’. You also see that “D” is transformed to “T” because they nearly sound the same.

If you think you have found a regularity you should always take your word list and grep for the corresponding regular expression you want to make a transformation rule for. An example: If you come to the idea that all English words ending on ‘ough’ sound like ‘AF’ at the end because you think of ‘enough’ and ‘tough’. If you then grep for the corresponding regular expression by ‘grep -i ough$ wordlist’ you will see that the rule you wanted to set up is not correct because the rule doesn’t fit to words like ‘although’ or ‘bough’. So you have to define your rule more precisely or you have to set up exceptions if the number of words that differ from the desired rule is not too big.

Don’t forget about follow-up rules which can help in many cases but which also can lead to confusion and unwanted side effects. It’s also important to write exceptions in front of the more general rules (‘GH’ before ‘G’ etc.).

If you think you have set up a number of rules that may produce some good results try them out! If you run Aspell as ‘aspell --lang=your_language pipe’ you get a prompt at which you can type in words. If you just type words Aspell checks them and eventually makes suggestions if they are misspelled. If you type in $$Sw word you will see the phonetic transformation and you can test out if your work does what you want.

Another good way to check that changes you make to your rules don’t have any bad side effects is to create another list from your word list which contains not only the word of the word list but also the corresponding phonetic version of this word on the same line. If you do this once before the change and once after the change you can make a diff (see ‘man diff’) to see what really changed. To do this use the command ‘aspell --lang=your_language soundslike’. In this mode Aspell will output the the original word and then its soundslike separated by a tab character for each word you give it. If you are interested in seeing how the algorithm works you can download a set of useful programs from http://members.xoom.com/maccy/spell/phonet-utils.tar.gz. This includes a program that produces a list as mentioned above and another program which illustrates how the algorithm works. It uses the same transformation table as Aspell and so it helps a lot during the process of creating a phonetic transformation table for Aspell.

During your work you should write down your basic ideas so that other people are able to understand what you did (and you still know about it after a few weeks). The English table has a huge documentation appended as an example.

Now you can start experimenting with all the things you just read and perhaps set up a nice phonetic transformation table for your language to help Aspell to come up with the best correction suggestions ever seen also for your language. Take a look at the Aspell homepage to see if there is already a transformation table for your language. If there is one you might also take a look at it to see if it could be improved.

If you think that this section helped you or if you think that this is just a waste of time you can send any feedback to bjoern.jacke@gmx.de.