*** DRAFT ***

The Spellfix1 Virtual Table

This spellfix1 virtual table can be used to search a large vocabulary for close matches. For example, spellfix1 can be used to suggest corrections to misspelled words. Or, it could be used with FTS4 to do full-text search using potentially misspelled words.

The implementation for the spellfix1 virtual table is held in the SQLite source tree in the miscellaneous extensions folder and in particular in the file ext/misc/spellfix1.c. The spellfix1 virtual table is not included in the SQLite amalgamation and is not a part of any standard SQLite build. It is a loadable extension.

Once the spellfix1 extension is loaded, an instance of the spellfix1 virtual table is created like this:

CREATE VIRTUAL TABLE demo USING spellfix1;

The "spellfix1" term is the name of the spellfix module and must be entered as shown. The "demo" term is the name of the virtual table you will be creating and can be altered to suit the needs of your application. The virtual table is initially empty. In order for the virtual table to be useful, you will need to populate it with your vocabulary. Suppose you have a list of words in a table named "big_vocabulary". Then do this:

INSERT INTO demo(word) SELECT word FROM big_vocabulary;

If you intend to use this virtual table in cooperation with an FTS4 table (for spelling correctly of search terms) then you might extract the vocabulary using an fts4aux table:

INSERT INTO demo(word) SELECT term FROM search_aux WHERE col='*';

You can also provide the virtual table with a "rank" for each word. The "rank" is an estimate of how common the word is. Larger numbers mean the word is more common. If you omit the rank when populating the table, then a rank of 1 is assumed. But if you have rank information, you can supply it and the virtual table will show a slight preference for selecting more commonly used terms. To populate the rank from an fts4aux table "search_aux" do something like this:

INSERT INTO demo(word,rank)
   SELECT term, documents FROM search_aux WHERE col='*';

To query the virtual table, include a MATCH operator in the WHERE clause. For example:

SELECT word FROM demo WHERE word MATCH 'kennasaw';

Using a dataset of American place names (derived from http://geonames.usgs.gov/domestic/download_data.htm) the query above returns 20 results beginning with:

kennesaw
kenosha
kenesaw
kenaga
keanak

If you append the character '*' to the end of the pattern, then a prefix search is performed. For example:

SELECT word FROM demo WHERE word MATCH 'kennes*';

Yields 20 results beginning with:

kennesaw
kennestone
kenneson
kenneys
keanes
keenes

Search Refinements

By default, the spellfix1 table returns no more than 20 results. (It might return less than 20 if there were fewer good matches.) You can change the upper bound on the number of returned rows by adding a "top=N" term to the WHERE clause of your query, where N is the new maximum. For example, to see the 5 best matches:

SELECT word FROM demo WHERE word MATCH 'kennes*' AND top=5;

Each entry in the spellfix1 virtual table is associated with a a particular language, identified by the integer "langid" column. The default langid is 0 and if no other actions are taken, the entire vocabulary is a part of the 0 language. But if your application needs to operate in multiple languages, then you can specify different vocabulary items for each language by specifying the langid field when populating the table. For example:

INSERT INTO demo(word,langid) SELECT word, 0 FROM en_vocabulary;
INSERT INTO demo(word,langid) SELECT word, 1 FROM de_vocabulary;
INSERT INTO demo(word,langid) SELECT word, 2 FROM fr_vocabulary;
INSERT INTO demo(word,langid) SELECT word, 3 FROM ru_vocabulary;
INSERT INTO demo(word,langid) SELECT word, 4 FROM cn_vocabulary;

After the virtual table has been populated with items from multiple languages, specify the language of interest using a "langid=N" term in the WHERE clause of the query:

SELECT word FROM demo WHERE word MATCH 'hildes*' AND langid=1;

Note that if you do not include the "langid=N" term in the WHERE clause, the search will be against language 0 (English in the example above.) All spellfix1 searches are against a single language id. There is no way to search all languages at once.

Virtual Table Details

Each row in the spellfix1 virtual table has a unique rowid with seven columns plus five extra hidden columns. The columns are as follows:

rowid

A unique integer number associated with each vocabulary item in the table. This can be used as a foreign key on other tables in the database.

word

The text of the word that matches the pattern. Both word and pattern can contains unicode characters and can be mixed case.

rank

This is the rank of the word, as specified in the original INSERT statement.

distance

This is an edit distance or Levensthein distance going from the pattern to the word.

langid

This is the language-id of the word. All queries are against a single language-id, which defaults to 0. For any given query this value is the same on all rows.

score

The score is a combination of rank and distance. The idea is that a lower score is better. The virtual table attempts to find words with the lowest score and by default (unless overridden by ORDER BY) returns results in order of increasing score.

matchlen

In a prefix search, the matchlen is the number of characters in the string that match against the prefix. For a non-prefix search, this is the same as length(word).

phonehash

This column shows the phonetic hash prefix that was used to restrict the search. For any given query, this column should be the same for every row. This information is available for diagnostic purposes and is not normally considered useful in real applications.

top

(HIDDEN) For any query, this value is the same on all rows. It is an integer which is the maximum number of rows that will be output. The actually number of rows output might be less than this number, but it will never be greater. The default value for top is 20, but that can be changed for each query by including a term of the form "top=N" in the WHERE clause of the query.

scope

(HIDDEN) For any query, this value is the same on all rows. The scope is a measure of how widely the virtual table looks for matching words. Smaller values of scope cause a broader search. The scope is normally chosen automatically and is capped at 4. Applications can change the scope by including a term of the form "scope=N" in the WHERE clause of the query. Increasing the scope will make the query run faster, but will reduce the possible corrections.

srchcnt

(HIDDEN) For any query, this value is the same on all rows. This value is an integer which is the number of of words examined using the edit-distance algorithm to find the top matches that are ultimately displayed. This value is for diagnostic use only.

soundslike

(HIDDEN) When inserting vocabulary entries, this field can be set to a spelling that matches what the word sounds like. See the DEALING WITH UNUSUAL AND DIFFICULT SPELLINGS section below for details.

command

(HIDDEN) The value of the "command" column is always NULL. However, applications can insert special strings into the "command" column in order to provoke certain behaviors in the spellfix1 virtual table. For example, inserting the string 'reset' into the "command" column will cause the virtual table to reread its edit distance weights (if there are any).

Algorithm

The spellfix1 virtual table creates a single shadow table named "%_vocab" (where the % is replaced by the name of the virtual table; Ex: "demo_vocab" for the "demo" virtual table). the shadow table contains the following columns:

id

The unique id (INTEGER PRIMARY KEY)

rank

The rank of word.

langid

The language id for this entry.

word

The original UTF8 text of the vocabulary word

k1

The word transliterated into lower-case ASCII. There is a standard table of mappings from non-ASCII characters into ASCII. Examples: "æ" -> "ae", "þ" -> "th", "ß" -> "ss", "á" -> "a", ... The accessory function spellfix1_translit(X) will do the non-ASCII to ASCII mapping. The built-in lower(X) function will convert to lower-case. Thus: k1 = lower(spellfix1_translit(word)).

k2

This field holds a phonetic code derived from k1. Letters that have similar sounds are mapped into the same symbol. For example, all vowels and vowel clusters become the single symbol "A". And the letters "p", "b", "f", and "v" all become "B". All nasal sounds are represented as "N". And so forth. The mapping is base on ideas found in Soundex, Metaphone, and other long-standing phonetic matching systems. This key can be generated by the function spellfix1_phonehash(X). Hence: k2 = spellfix1_phonehash(k1)

There is also a function for computing the Wagner edit distance or the Levenshtein distance between a pattern and a word. This function is exposed as spellfix1_editdist(X,Y). The edit distance function returns the "cost" of converting X into Y. Some transformations cost more than others. Changing one vowel into a different vowel, for example is relatively cheap, as is doubling a constant, or omitting the second character of a double-constant. Other transformations or more expensive. The idea is that the edit distance function returns a low cost for words that are similar and a higher cost for words that are further apart. In this implementation, the maximum cost of any single-character edit (delete, insert, or substitute) is 100, with lower costs for some edits (such as transforming vowels).

The "score" for a comparison is the edit distance between the pattern and the word, adjusted down by the base-2 logarithm of the word rank. For example, a match with distance 100 but rank 1000 would have a score of 122 (= 100 - log2(1000) + 32) where as a match with distance 100 with a rank of 1 would have a score of 131 (100 - log2(1) + 32). (NB: The constant 32 is added to each score to keep it from going negative in case the edit distance is zero.) In this way, frequently used words get a slightly lower cost which tends to move them toward the top of the list of alternative spellings.

A straightforward implementation of a spelling corrector would be to compare the search term against every word in the vocabulary and select the 20 with the lowest scores. However, there will typically be hundreds of thousands or millions of words in the vocabulary, and so this approach is not fast enough.

Suppose the term that is being spell-corrected is X. To limit the search space, X is converted to a k2-like key using the equivalent of:

   key = spellfix1_phonehash(lower(spellfix1_translit(X)))

This key is then limited to "scope" characters. The default scope value is 4, but an alternative scope can be specified using the "scope=N" term in the WHERE clause. After the key has been truncated, the edit distance is run against every term in the vocabulary that has a k2 value that begins with the abbreviated key.

For example, suppose the input word is "Paskagula". The phonetic key is "BACACALA" which is then truncated to 4 characters "BACA". The edit distance is then run on the 4980 entries (out of 272,597 entries total) of the vocabulary whose k2 values begin with BACA, yielding "Pascagoula" as the best match.

Only terms of the vocabulary with a matching langid are searched. Hence, the same table can contain entries from multiple languages and only the requested language will be used. The default langid is 0.

Configurable Edit Distance

The built-in Wagner edit-distance function with fixed weights can be replaced by the editdist3() edit-distance function with application-defined weights and support for unicode, by specifying the "edit_cost_table=TABLENAME" parameter to the spellfix1 module when the virtual table is created. For example:

CREATE VIRTUAL TABLE demo2 USING spellfix1(edit_cost_table=APPCOST);

The editdist3() edit-distance function can also be selected or deselected at run-time by inserting an appropriate string into the "command" column of the virtual table:

INSERT INTO demo2(command) VALUES('edit_cost_table=APPCOST');

In the examples above, the APPCOST table would be interrogated to find the edit distance coefficients. It is the presence of the "edit_cost_table=" parameter to the spellfix1 module name that causes editdist3() to be used in place of the built-in edit distance function. If APPCOST is an empty string, then the built-in Wagner edit-distance function is used.

The edit distance coefficients are normally read from the APPCOST table once and there after stored in memory. Hence, run-time changes to the APPCOST table will not normally effect the edit distance results. However, inserting the special string 'reset' into the "command" column of the virtual table causes the edit distance coefficients to be reread the APPCOST table. Hence, applications should run a SQL statement similar to the following when changes to the APPCOST table occur:

INSERT INTO demo2(command) VALUES("reset");

Dealing With Unusual And Difficult Spellings

The algorithm above works quite well for most cases, but there are exceptions. These exceptions can be dealt with by making additional entries in the virtual table using the "soundslike" column.

For example, many words of Greek origin begin with letters "ps" where the "p" is silent. Ex: psalm, pseudonym, psoriasis, psyche. In another example, many Scottish surnames can be spelled with an initial "Mac" or "Mc". Thus, "MacKay" and "McKay" are both pronounced the same.

Accommodation can be made for words that are not spelled as they sound by making additional entries into the virtual table for the same word, but adding an alternative spelling in the "soundslike" column. For example, the canonical entry for "psalm" would be this:

  INSERT INTO demo(word) VALUES('psalm');

To enhance the ability to correct the spelling of "salm" into "psalm", make an addition entry like this:

  INSERT INTO demo(word,soundslike) VALUES('psalm','salm');

It is ok to make multiple entries for the same word as long as each entry has a different soundslike value. Note that if no soundslike value is specified, the soundslike defaults to the word itself.

Listed below are some cases where it might make sense to add additional soundslike entries. The specific entries will depend on the application and the target language.

Auxiliary Functions

The source code module that implements the spellfix1 virtual table also implements several SQL functions that might be useful to applications that employ spellfix1 or for testing or diagnostic work while developing applications that use spellfix1. The following auxiliary functions are available:

editdist3(P,W)
editdist3(P,W,L)
editdist3(T)

These routines provide direct access to the version of the Wagner edit-distance function that allows for application-defined weights on edit operations. The first two forms of this function compare pattern P against word W and return the edit distance. In the first function, the langid is assumed to be 0 and in the second, the langid is given by the L parameter. The third form of this function reloads edit distance coefficients from the table named by T.

spellfix1_editdist(P,W)

This routine provides access to the built-in Wagner edit-distance function that uses default, fixed costs. The value returned is the edit distance needed to transform W into P.

spellfix1_phonehash(X)

This routine constructs a phonetic hash of the pure ascii input word X and returns that hash. This routine is used internally by spellfix1 in order to transform the K1 column of the shadow table into the K2 column.

spellfix1_scriptcode(X)

Given an input string X, this routine attempts to determine the dominant script of that input and returns the ISO-15924 numeric code for that script. The current implementation understands the following scripts:
  • 215 - Latin
  • 220 - Cyrillic
  • 200 - Greek
Additional language codes might be added in future releases.

spellfix1_translit(X)

This routine transliterates unicode text into pure ascii, returning the pure ascii representation of the input text X. This is the function that is used internally to transform vocabulary words into the K1 column of the shadow table.

The editdist3 function

The editdist3 algorithm is a function that computes the minimum edit distance (a.k.a. the Levenshtein distance) between two input strings. The editdist3 algorithm is a configurable alternative to the default edit distance function of spellfix1. Features of editdist3 include:

The editdist3 COST table

To program the costs of editdist3, create a table such as the following:

CREATE TABLE editcost(
  iLang INT,   -- The language ID
  cFrom TEXT,  -- Convert text from this
  cTo   TEXT,  -- Convert text into this
  iCost INT    -- The cost of doing the conversion
);

The cost table can be named anything you want - it does not have to be called "editcost". And the table can contain additional columns. The only requirement is that the table must contain the four columns show above, with exactly the names shown.

The iLang column is a non-negative integer that identifies a set of costs appropriate for a particular language. The editdist3 function will only use a single iLang value for any given edit-distance computation. The default value is 0. It is recommended that applications that only need to use a single language always use iLang==0 for all entries.

The iCost column is the numeric cost of transforming cFrom into cTo. This value should be a non-negative integer, and should probably be less than 100. The default single-character insertion and deletion costs are 100 and the default single-character to single-character substitution cost is 150. A cost of 10000 or more is considered "infinite" and causes the rule to be ignored.

The cFrom and cTo columns show edit transformation strings. Either or both columns may contain more than one character. Or either column (but not both) may hold an empty string. When cFrom is empty, that is the cost of inserting cTo. When cTo is empty, that is the cost of deleting cFrom.

In the spellfix1 algorithm, cFrom is the text as the user entered it and cTo is the correctly spelled text as it exists in the database. The goal of the editdist3 algorithm is to determine how close the user-entered text is to the dictionary text.

There are three special-case entries in the cost table:

cFromcToMeaning
'''?'The default insertion cost
'?'''The default deletion cost
'?''?'The default substitution cost

If any of the special-case entries shows above are omitted, then the value of 100 is used for insertion and deletion and 150 is used for substitution. To disable the default insertion, deletion, and/or substitution set their respective cost to 10000 or more.

Other entries in the cost table specific transforms for particular characters. The cost of specific transforms should be less than the default costs, or else the default costs will take precedence and the specific transforms will never be used.

Some example, cost table entries:

INSERT INTO editcost(iLang, cFrom, cTo, iCost)
VALUES(0, 'a', 'ä', 5);

The rule above says that the letter "a" in user input can be matched against the letter "ä" in the dictionary with a penalty of 5.

INSERT INTO editcost(iLang, cFrom, cTo, iCost)
VALUES(0, 'ss', 'ß', 8);

The number of characters in cFrom and cTo do not need to be the same. The rule above says that "ss" on user input will match "ß" with a penalty of 8.

Experimenting with the editcost3() function

The spellfix1 virtual table uses editdist3 if the "edit_cost_table=TABLE" option is specified as an argument when the spellfix1 virtual table is created. But editdist3 can also be tested directly using the built-in "editdist3()" SQL function. The editdist3() SQL function has 3 forms:

  1. editdist3('TABLENAME');
  2. editdist3('string1', 'string2');
  3. editdist3('string1', 'string2', langid);

The first form loads the edit distance coefficients from a table called 'TABLENAME'. Any prior coefficients are discarded. So when experimenting with weights and the weight table changes, simply rerun the single-argument form of editdist3() to reload revised coefficients. Note that the edit distance weights used by the editdist3() SQL function are independent from the weights used by the spellfix1 virtual table.

The second and third forms return the computed edit distance between strings 'string1' and "string2'. In the second form, a language id of 0 is used. The language id is specified in the third form.

*** DRAFT ***