root / tags / 1.0.8 / xapian-core / docs / spelling.rst

.. Copyright (C) 2007 Olly Betts

==========================
Xapian Spelling Correction
==========================

.. contents:: Table of contents

Introduction
============

Xapian provides functionality which can suggest corrections for misspelled
words in queries, or in other situations where it might be useful.  The
suggestions are generated dynamically from the data that has been indexed, so
the correction facility isn't tied to particular languages, and it can suggest
proper nouns or specialist technical terms.

Algorithm
=========

A list of candidate words is generated by matching trigrams (groups of 3
adjacent characters) in the candidates against those in the misspelled
word.  As well as groups of adjacent characters, "starts" and "ends"
are generated with the first two and last two characters respectively
(e.g. "FISH" generates: "<start>FI", "FIS", "ISH", and "SH<end>").

This technique alone would missing many single-edit errors in two and three
character words, so we handle these specially as follows:

For a three character word (e.g. "ABC"), we generate trigrams for the two
transposed forms too ("BAC" and "ACB"), in addition to "<start>AB", "ABC",
and "BC<end>".

For a two character word (e.g. "AB"), we generate the special start and end
trigrams for the reversed form (i.e. "BA"), so the trigrams are "<start>AB",
"AB<end>", "<start>BA", and "BA<end>".

And for two, three, and four character words, we generate "bookend" bigrams
consisting of the prefix 'B' followed by the first and last letters.  This
allows us to handle transposition of the middle two characters of a four
letter word, substitution or deletion of the middle character of a three
letter word, or insertion in the middle of a two letter word.

Note that we don't attempt to suggest corrections for single character words
at all, since the suggestions are unlikely to be of good quality (we'd always
suggest the same correction for a given database, probably "a" for English).
We also don't currently attempt to suggest substitution corrections for two
character words, though this would perhaps be useful in some cases.

Those candidates with the better trigram matches are compared to the misspelled
word by calculating the "edit distance" - that's the smallest number of
operations required to turn one word into another.  The allowed operations
are: insert a character; delete a character; change a character to another;
transpose two adjacent characters.  The candidate with the smallest edit
distance is found, and if more than one word has the smallest edit distance,
that which occurs the most times is chosen.  If there's a tie of this too,
it's essentially arbitrary which is chosen.

The maximum edit distance to consider can be specified as an optional parameter
to Xapian::Database::get_spelling_suggestion().  If not specified, the default
is 2, which generally does a good job.  3 is also a reasonable choice in many
cases.  For most uses, 1 is probably too low, and 4 or more probably too high.

Unicode Support
---------------

Trigrams are generated at the byte level, but the edit distance calculation
currently works with Unicode characters, so get_spelling_suggestion() should
suggest suitable spelling corrections respecting the specified (or default)
edit distance threshold.

QueryParser Integration
=======================

If FLAG_SPELLING_CORRECTION is passed to QueryParser::parse_query() and
QueryParser::set_database() has been called, the QueryParser will look for
corrections for words in the query which aren't found in the database.

If a correction is found, then a modified version of the query string will be
generated which can be obtained by calling
QueryParser::get_corrected_query_string().  However, the original query string
will still be parsed, since you'll often want to ask the user "Did you mean:
[...] ?" - if you want to automatically use the corrected form, just call
QueryParser::parse_query() on it.

Current Limitations
===================

Exactness
---------

Because Xapian only tests the edit distance for terms which match
well (or at all!) on trigrams, it may not always suggest the same answer that
would be found if all possible words were checked using the edit distance
algorithm.  However, the best answer will usually be found, and an exhaustive
search would be prohibitively expensive for many uses.

Backend Support
---------------

Currently spelling correction is only supported for flint databases.  It
works with a single database or multiple databases (use
Database::add_database() as usual).  We've no plans to support it for the
deprecated Quartz backend, nor for InMemory, but we do intend to support it for
the remote backend in the future.

Prefixed Terms
--------------

Currently spelling correction ignores prefixed terms.

Omega
-----

Spelling correction hasn't yet been integrated into Omega.

QueryParser changed word locations
----------------------------------

The QueryParser doesn't currently report the locations of changed words in
the query string, so it's a bit fiddly to mark up the altered words specially
in HTML output, for example.

References
==========

The algorithm used to calculate the edit distance is based on that described in
the paper "An extension of Ukkonen's enhanced dynamic programming ASM
algorithm" by Hal Berghel, University of Arkansas, and David Roach, Acxiom
Corporation.  It's available online at:
http://berghel.net/publications/asm/asm.php