Spell correction

Spell correction, also known as:

  • Auto correction
  • Text correction
  • Fixing spelling errors
  • Typo tolerance
  • "Did you mean?"

and so on, is a software functionality that suggests alternatives to or makes automatic corrections of the text you have typed in. The concept of correcting typed text dates back to the 1960s when computer scientist Warren Teitelman, who also invented the "undo" command, introduced a philosophy of computing called D.W.I.M., or "Do What I Mean." Instead of programming computers to accept only perfectly formatted instructions, Teitelman argued that they should be programmed to recognize obvious mistakes.

The first well-known product to provide spell correction functionality was Microsoft Word 6.0, released in 1993.

How it works

There are a few ways spell correction can be done, but it's important to note that there is no purely programmatic way to convert your mistyped "ipone" into "iphone" with decent quality. Mostly, there has to be a dataset the system is based on. The dataset can be:

  • A dictionary of properly spelled words, which in turn can be:
    • Based on your real data. The idea here is that, for the most part, the spelling in the dictionary made up of your data is correct, and the system tries to find a word that is most similar to the typed word (we'll discuss how this can be done with Manticore shortly).
    • Or it can be based on an external dictionary unrelated to your data. The issue that may arise here is that your data and the external dictionary can be too different: some words may be missing in the dictionary, while others may be missing in your data.
  • Not just dictionary-based, but also context-aware, e.g., "white ber" would be corrected to "white bear," while "dark ber" would be corrected to "dark beer." The context might not just be a neighboring word in your query, but also your location, time of day, the current sentence's grammar (to change "there" to "their" or not), your search history, and virtually any other factors that can affect your intent.
  • Another classic approach is to use previous search queries as the dataset for spell correction. This is even more utilized in autocomplete functionality but makes sense for autocorrect too. The idea is that users are mostly right with spelling, so we can use words from their search history as a source of truth, even if we don't have the words in our documents or use an external dictionary. Context-awareness is also possible here.

Manticore provides the commands CALL QSUGGEST and CALL SUGGEST that can be used for automatic spell correction purposes.

CALL QSUGGEST, CALL SUGGEST

Both commands are available via SQL only, and the general syntax is:

CALL QSUGGEST(word, table [,options])
CALL SUGGEST(word, table [,options])

options: N as option_name[, M as another_option, ...]

These commands provide all suggestions from the dictionary for a given word. They work only on tables with infixing enabled and dict=keywords. They return the suggested keywords, Levenshtein distance between the suggested and original keywords, and the document statistics of the suggested keyword.

If the first parameter contains multiple words, then:

  • CALL QSUGGEST will return suggestions only for the last word, ignoring the rest.
  • CALL SUGGEST will return suggestions only for the first word.

That's the only difference between them. Several options are supported for customization:

Option Description Default
limit Returns N top matches 5
max_edits Keeps only dictionary words with a Levenshtein distance less than or equal to N 4
result_stats Provides Levenshtein distance and document count of the found words 1 (enabled)
delta_len Keeps only dictionary words with a length difference less than N 3
max_matches Number of matches to keep 25
reject Rejected words are matches that are not better than those already in the match queue. They are put in a rejected queue that gets reset in case one actually can go in the match queue. This parameter defines the size of the rejected queue (as reject*max(max_matched,limit)). If the rejected queue is filled, the engine stops looking for potential matches 4
result_line alternate mode to display the data by returning all suggests, distances and docs each per one row 0
non_char do not skip dictionary words with non alphabet symbols 0 (skip such words)
sentence Returns the original sentence along with the last word replaced by the matched one. 0 (do not return the full sentence)

To show how it works, let's create a table and add a few documents to it.

create table products(title text) min_infix_len='2';
insert into products values (0,'Crossbody Bag with Tassel'), (0,'microfiber sheet set'), (0,'Pet Hair Remover Glove');
Single word example

As you can see, the mistyped word "crossbUdy" gets corrected to "crossbody". By default, CALL SUGGEST/QSUGGEST return:

  • distance - the Levenshtein distance which means how many edits they had to make to convert the given word to the suggestion
  • docs - number of documents containing the suggested word

To disable the display of these statistics, you can use the option 0 as result_stats.

‹›
  • Example
Example
📋
call suggest('crossbudy', 'products');
‹›
Response
+-----------+----------+------+
| suggest   | distance | docs |
+-----------+----------+------+
| crossbody | 1        | 1    |
+-----------+----------+------+
CALL SUGGEST takes only the first word

If the first parameter is not a single word, but multiple, then CALL SUGGEST will return suggestions only for the first word.

‹›
  • Example
Example
📋
call suggest('bagg with tasel', 'products');
‹›
Response
+---------+----------+------+
| suggest | distance | docs |
+---------+----------+------+
| bag     | 1        | 1    |
+---------+----------+------+
CALL QSUGGEST takes only the last word

If the first parameter is not a single word, but multiple, then CALL SUGGEST will return suggestions only for the last word.

‹›
  • Example
Example
📋
CALL QSUGGEST('bagg with tasel', 'products');
‹›
Response
+---------+----------+------+
| suggest | distance | docs |
+---------+----------+------+
| tassel  | 1        | 1    |
+---------+----------+------+

Adding 1 as sentence makes CALL QSUGGEST return the entire sentence with the last word corrected.

‹›
  • Example
Example
📋
CALL QSUGGEST('bag with tasel', 'products', 1 as sentence);
‹›
Response
+-------------------+----------+------+
| suggest           | distance | docs |
+-------------------+----------+------+
| bag with tassel   | 1        | 1    |
+-------------------+----------+------+
Different display mode

The 1 as result_line option changes the way the suggestions are displayed in the output. Instead of showing each suggestion in a separate row, it displays all suggestions, distances, and docs in a single row. Here's an example to demonstrate this:

Interactive course

This interactive course demonstrates online how the spell correction feature works on a web page and experiment with different examples.

Typical flow with Manticore and a database

Query cache

Query cache stores compressed result sets in memory and reuses them for subsequent queries when possible. You can configure it using the following directives:

  • qcache_max_bytes, a limit on the RAM usage for cached query storage. Defaults to 16 MB. Setting qcache_max_bytes to 0 completely disables the query cache.
  • qcache_thresh_msec, the minimum wall query time to cache. Queries that complete faster than this will not be cached. Defaults to 3000 msec, or 3 seconds.
  • qcache_ttl_sec, cached entry TTL, or time to live. Queries will stay cached for this duration. Defaults to 60 seconds, or 1 minute.

These settings can be changed on the fly using the SET GLOBAL statement:

mysql> SET GLOBAL qcache_max_bytes=128000000;

These changes are applied immediately, and cached result sets that no longer satisfy the constraints are immediately discarded. When reducing the cache size on the fly, MRU (most recently used) result sets win.

Query cache operates as follows. When enabled, every full-text search result is completely stored in memory. This occurs after full-text matching, filtering, and ranking, so essentially we store total_found docid,weight} pairs. Compressed matches can consume anywhere from 2 bytes to 12 bytes per match on average, mostly depending on the deltas between subsequent docids. Once the query is complete, we check the wall time and size thresholds, and either save the compressed result set for reuse or discard it.

Note that the query cache's impact on RAM is not limited byqcache_max_bytes! If you run, for example, 10 concurrent queries, each matching up to 1M matches (after filters), then the peak temporary RAM usage will be in the range of 40 MB to 240 MB, even if the queries are fast enough and don't get cached.

Queries can use cache when the table, full-text query (i.e.,MATCH() contents), and ranker all match, and filters are compatible. This means:

  • The full-text part within MATCH() must be a bytewise match. Add a single extra space, and it's now a different query as far as the query cache is concerned.
  • The ranker (and its parameters, if any, for user-defined rankers) must be a bytewise match.
  • The filters must be a superset of the original filters. You can add extra filters and still hit the cache. (In this case, the extra filters will be applied to the cached result.) But if you remove one, that will be a new query again.

Cache entries expire with TTL and also get invalidated on table rotation, or on TRUNCATE, or on ATTACH. Note that currently, entries are not invalidated on arbitrary RT table writes! So a cached query might return older results for the duration of its TTL.

You can inspect the current cache status with SHOW STATUS through the qcache_XXX variables:

mysql> SHOW STATUS LIKE 'qcache%';
+-----------------------+----------+
| Counter               | Value    |
+-----------------------+----------+
| qcache_max_bytes      | 16777216 |
| qcache_thresh_msec    | 3000     |
| qcache_ttl_sec        | 60       |
| qcache_cached_queries | 0        |
| qcache_used_bytes     | 0        |
| qcache_hits           | 0        |
+-----------------------+----------+
6 rows in set (0.00 sec)

Collations

Collations primarily impact string attribute comparisons. They define both the character set encoding and the strategy Manticore employs for comparing strings when performing ORDER BY or GROUP BY with a string attribute involved.

String attributes are stored as-is during indexing, and no character set or language information is attached to them. This is fine as long as Manticore only needs to store and return the strings to the calling application verbatim. However, when you ask Manticore to sort by a string value, the request immediately becomes ambiguous.

First, single-byte (ASCII, ISO-8859-1, or Windows-1251) strings need to be processed differently than UTF-8 strings, which may encode each character with a variable number of bytes. Thus, we need to know the character set type to properly interpret the raw bytes as meaningful characters.

Second, we also need to know the language-specific string sorting rules. For example, when sorting according to US rules in the en_US locale, the accented character ï (small letter i with diaeresis) should be placed somewhere after z. However, when sorting with French rules and the fr_FR locale in mind, it should be placed between i and j. Some other set of rules might choose to ignore accents altogether, allowing ï and i to be mixed arbitrarily.

Third, in some cases, we may require case-sensitive sorting, while in others, case-insensitive sorting is needed.

Collations encapsulate all of the following: the character set, the language rules, and the case sensitivity. Manticore currently provides four collations:

  1. libc_ci
  2. libc_cs
  3. utf8_general_ci
  4. binary

The first two collations rely on several standard C library (libc) calls and can thus support any locale installed on your system. They provide case-insensitive (_ci) and case-sensitive (_cs) comparisons, respectively. By default, they use the C locale, effectively resorting to bytewise comparisons. To change that, you need to specify a different available locale using the collation_libc_locale directive. The list of locales available on your system can usually be obtained with the locale command:

$ locale -a
C
en_AG
en_AU.utf8
en_BW.utf8
en_CA.utf8
en_DK.utf8
en_GB.utf8
en_HK.utf8
en_IE.utf8
en_IN
en_NG
en_NZ.utf8
en_PH.utf8
en_SG.utf8
en_US.utf8
en_ZA.utf8
en_ZW.utf8
es_ES
fr_FR
POSIX
ru_RU.utf8
ru_UA.utf8

The specific list of system locales may vary. Consult your OS documentation to install additional needed locales.

utf8_general_ci and binary locales are built-in into Manticore. The first one is a generic collation for UTF-8 data (without any so-called language tailoring); it should behave similarly to the utf8_general_ci collation in MySQL. The second one is a simple bytewise comparison.

Collation can be overridden via SQL on a per-session basis using the SET collation_connection statement. All subsequent SQL queries will use this collation. Otherwise, all queries will use the server default collation or as specified in the collation_server configuration directive. Manticore currently defaults to the libc_ci collation.

Collations affect all string attribute comparisons, including those within ORDER BY and GROUP BY, so differently ordered or grouped results can be returned depending on the collation chosen. Note that collations don't affect full-text searching; for that, use the charset_table.