U.S. patent application number 13/250834 was filed with the patent office on 2013-12-26 for method of spell-checking search queries.
This patent application is currently assigned to GOOGLE INC.. The applicant listed for this patent is Noam SHAZEER. Invention is credited to Noam SHAZEER.
Application Number | 20130346434 13/250834 |
Document ID | / |
Family ID | 37863976 |
Filed Date | 2013-12-26 |
United States Patent
Application |
20130346434 |
Kind Code |
A1 |
SHAZEER; Noam |
December 26, 2013 |
METHOD OF SPELL-CHECKING SEARCH QUERIES
Abstract
A computer-implemented method for determining whether a target
text-string is correctly spelled is provided. The target
text-string is compared to a corpus to determine a set of contexts
which each include an occurrence of the target text-string. Using
heuristics, each context of the set is characterized based on
occurrences in the corpus of the target text-string and a reference
text-string. Contexts are characterized as including a correct
spelling of the target text-string, an incorrect spelling of the
reference text-string, or including an indeterminate usage of the
target text-string. A likelihood that the target text-string is a
misspelling of the reference text-string is computed as a function
of the quantity of contexts including a correct spelling of the
target text-string and the quantity of contexts including an
incorrect spelling of a reference text-string. In one application,
the target text-string is received in a search query, the search
executed following a spell-check.
Inventors: |
SHAZEER; Noam; (Mountain
View, CA) |
|
Applicant: |
Name |
City |
State |
Country |
Type |
SHAZEER; Noam |
Mountain View |
CA |
US |
|
|
Assignee: |
GOOGLE INC.
Mountain View
CA
|
Family ID: |
37863976 |
Appl. No.: |
13/250834 |
Filed: |
September 30, 2011 |
Related U.S. Patent Documents
|
|
|
|
|
|
Application
Number |
Filing Date |
Patent Number |
|
|
11670885 |
Feb 2, 2007 |
8051374 |
|
|
13250834 |
|
|
|
|
10119375 |
Apr 9, 2002 |
7194684 |
|
|
11670885 |
|
|
|
|
Current U.S.
Class: |
707/759 ;
707/E17.005; 707/E17.017 |
Current CPC
Class: |
Y10S 707/99936 20130101;
Y10S 707/99933 20130101; G06F 40/232 20200101 |
Class at
Publication: |
707/759 ;
707/E17.005; 707/E17.017 |
International
Class: |
G06F 17/30 20060101
G06F017/30 |
Claims
1-36. (canceled)
37. A method comprising: receiving a search query that includes a
query term; identifying, from a corpus of documents, text patterns
that each include the query term occurring adjacent to one or more
other query terms; determining a first quantity of occurrences of
the text patterns that each include the query term occurring
adjacent to the one or more other terms, in the corpus of
documents; determining a second quantity of occurrences of text
patterns that each include a heterographic homophone of the query
term occurring adjacent to the one or more other terms, in the
corpus of documents; and determining, by one or more computers,
whether to revise the received search query to include the
heterographic homophone of the query term, based on comparing the
first quantity and the second quantity.
38-63. (canceled)
64. The method of claim 37, wherein determining whether to revise
is further based on determining that the first quantity exceeds a
threshold value.
65. The method of claim 37, wherein determining whether to revise
is further based on determining that the first quantity does not
exceed the threshold value, and that the second quantity exceeds
the threshold value.
66. The method of claim 37, wherein determining whether to revise
comprises determining not to revise when a ratio of the first
quantity to the second quantity exceeds a threshold value.
67. The method of claim 37, wherein determining whether to revise
comprises determining to revise when the second quantity exceeds a
first threshold value, and a ratio of the second quantity to the
first quantity exceeds a second threshold value.
68. The method of claim 37, wherein determining whether to revise
comprises determining not to revise when a ratio of the first
quantity to a sum of the first quantity and the second quantity
exceeds a threshold value.
69. The method of claim 37, wherein determining whether to revise
comprises determining not to revise when a ratio of the second
quantity to a sum of the first quantity and the second quantity is
less than a threshold value.
70. The method of claim 37, wherein the corpus of documents is a
first corpus of documents, and wherein the method further
comprises: determining a third quantity of occurrences of the
received search query in a second corpus of documents, wherein the
first corpus of documents includes less misspellings than the
second corpus of documents, wherein the received search query is
selected when a ratio of the first quantity to the third quantity
exceeds a threshold value.
71. The method of claim 37, wherein the corpus of documents is a
first corpus of documents, and wherein the method further
comprises: determining a third quantity of occurrences of the
received search query in a second corpus of documents, wherein the
first corpus of documents includes less misspellings than the
second corpus of documents, determining a fourth quantity of
occurrences of the modified search query in the second corpus of
documents, wherein the received search query is selected when a
ratio of the first quantity to the third quantity exceeds a ratio
of the second quantity to the fourth quantity.
72. A system comprising: one or more computers and one or more
storage devices storing instructions that are operable, when
executed by the one or more computers, to cause the one or more
computers to perform operations comprising: receiving a search
query that includes a query term; identifying, from a corpus of
documents, text patterns that each include the query term occurring
adjacent to one or more other query terms; determining a first
quantity of occurrences of the text patterns that each include the
query term occurring adjacent to the one or more other terms, in
the corpus of documents; determining a second quantity of
occurrences of text patterns that each include a heterographic
homophone of the query term occurring adjacent to the one or more
other terms, in the corpus of documents; and determining, by one or
more computers, whether to revise the received search query to
include the heterographic homophone of the query term, based on
comparing the first quantity and the second quantity.
73. The system of claim 72, wherein determining whether to revise
is further based on determining that the first quantity exceeds a
threshold value.
74. The system of claim 72, wherein determining whether to revise
is further based on determining that the first quantity does not
exceed the threshold value, and the second quantity exceeds a
threshold value.
75. The system of claim 72, wherein determining whether to revise
comprises determining not to revise the received search query when
a ratio of the first quantity to the second quantity exceeds a
threshold value.
76. The system of claim 72, wherein determining whether to revise
comprises determining not to revise the search query when the first
quantity exceeds a first threshold value, and a ratio of the first
quantity to the second quantity exceeds a second threshold
value.
77. The system of claim 72, wherein determining whether to revise
comprises determining not to revise the received search query when
a ratio of the first quantity to a sum of the first quantity and
the second quantity exceeds a threshold value.
78. The system of claim 72, wherein determining whether to revise
comprises determining not to revise when a ratio of the second
quantity to a sum of the first quantity and the second quantity is
less than a threshold value.
79. The system of claim 72, wherein the corpus of documents is a
first corpus of documents, and the operations further comprise:
accessing a second corpus of documents; and determining a third
quantity of occurrences of the received search query in the second
corpus of documents, where the first corpus of documents includes
less misspellings than the second corpus of documents, wherein the
received search query is selected when a ratio of the first
quantity to the third quantity exceeds a threshold value.
80. The system of claim 72, wherein the corpus of documents is a
first corpus of documents, and the operations further comprise:
determining a third quantity of occurrences of the received search
query in a second corpus of documents, wherein the first corpus of
documents includes less misspellings than the second corpus of
documents, determining a fourth quantity of occurrences of the
modified search query in the second corpus of documents, wherein
the received search query is selected when a ratio of the first
quantity to the third quantity exceeds a ratio of the second
quantity to the fourth quantity.
81. A non-transitory computer-readable storage device storing
software comprising instructions executable by one or more
computers which, upon such execution, cause the one or more
computers to perform operations comprising: receiving a search
query that includes a query term; identifying, from a corpus of
documents, text patterns that each include the query term occurring
adjacent to one or more other query terms; determining a first
quantity of occurrences of the text patterns that each include the
query term occurring adjacent to the one or more other terms, in
the corpus of documents; determining a second quantity of
occurrences of text patterns that each include a heterographic
homophone of the query term occurring adjacent to the one or more
other terms, in the corpus of documents; and determining, by one or
more computers, whether to revise the received search query to
include the heterographic homophone of the query term, based on
comparing the first quantity and the second quantity.
82. The device of claim 81, wherein determining whether to revise
is further based on determining that the first quantity does not
exceed the threshold value, and the second quantity exceeds a
threshold value.
83. The device of claim 81, wherein determining whether to revise
comprises determining not to revise when a ratio of the first
quantity to the second quantity exceeds a threshold value.
84. The device of claim 81, wherein determining whether to revise
comprises determining not to revise when the first quantity exceeds
a first threshold value, and a ratio of the first quantity to the
second quantity exceeds a second threshold value.
85. The device of claim 81, wherein determining whether to revise
comprises determining not to revise when a ratio of the first
quantity to a sum of the first quantity and the second quantity
exceeds a threshold value.
86. The device of claim 81, wherein determining whether to revise
comprises determining not to revise when a ratio of the first
quantity to a sum of the first quantity and the second quantity
exceeds a threshold value.
87. The device of claim 81, wherein determining whether to revise
comprises determining not to revise when a ratio of the second
quantity to a sum of the first quantity and the second quantity is
less than a threshold value.
88. The device of claim 81, the operations further comprising:
accessing a second corpus of documents; and determining a third
quantity of occurrences of the received search query in a second
corpus of documents, where the first corpus of documents includes
less misspellings than the second corpus of documents, where the
received search query is selected when a ratio of the first
quantity to the third quantity exceeds a threshold value.
89. The device of claim 81, wherein the corpus of documents is a
first corpus of documents, and wherein the operations further
comprise: determining a third quantity of occurrences of the
received search query in a second corpus of documents, where the
first corpus of documents includes less misspellings than the
second corpus of documents; and determining a fourth quantity of
occurrences of the modified search query in the second corpus of
documents, wherein the received search query is selected when a
ratio of the first quantity to the third quantity exceeds a ratio
of the second quantity to the fourth quantity.
90. The method of claim 37, wherein determining whether to revise
is based on a ratio comprising the first quantity and the second
quantity.
Description
CROSS-REFERENCE TO RELATED APPLICATIONS
[0001] This relates to pending U.S. patent application Ser. No.
______ (filed on ______), which is in turn a continuation of U.S.
patent application Ser. No. 09/004,827, entitled "Method for Node
Ranking in a Linked Database," filed on Jan. 9, 1998, now U.S. Pat.
No. 6,285,999, which claims priority to U.S. Provisional Patent
Application Ser. No. 60/035,205 filed Jan. 10, 1997, priority to
which is claimed under 35 U.S.C. .sctn.120 for any common subject
matter.
FIELD OF THE INVENTION
[0002] The present invention generally relates to retrieving
information from a data communication network and, more
particularly, to techniques for computer-implemented spell checking
of search engine query text strings.
BACKGROUND OF THE INVENTION
[0003] The World Wide Web (or "Web") contains a vast amount of
information in the form of hyperlinked documents (e.g., web pages)
loosely-organized and accessed through a data communication network
(or "Internet"). One of the reasons for the virtually explosive
growth in the number of hyperlinked documents on the Web is that
just about anyone can upload hyperlinked documents, which can
include links to other hyperlinked documents. The unstructured
nature and sheer volume of web pages available via the Internet
makes it difficult to efficiently find and navigate through related
information while avoiding unrelated information.
[0004] One conventional way to cull information on a computer
network (e.g., the Internet) is through use of a search engine. A
user typically begins a search for relevant information using a
search engine. A search engine attempts to return relevant
information in response to a request from a user. This request
usually comes in the form of a query (e.g., a set of words that are
related to a desired topic). Search engines typically return a
number of links to web pages, with a brief description of those
pages. Because the vast number of pages on the Web, ensuring that
the returned pages are relevant to the topic the user had in mind
is a central problem in web searching. Possibly the simplest and
most prevalent way of searching the web is to search for web pages
which have a relation to, or containing, all or many of the words
included in the query. Such a method is typically referred to as
text-based searching. Text-based searching over the Web can be
notoriously imprecise and several problems can arise in the
process.
[0005] The process of searching the Internet for narrowly-defined
relevant information is akin to finding a "needle" of relevant
information in a "haystack" of all the possible information
available through the data network. The efficiency of the search
process is greatly dependant on the quality of the search. Often a
large number of web pages match a user's query. Typically,
presentation of query results are ranked according to a predefined
method or criteria thereby directing a user to what is believed to
be the most-relevant information first. Poor quality queries tend
to misdirect the search process, interfere with ranking algorithms
and generally, produce poorer search results. In the aggregate,
inefficient Internet search methods tend to slow the data network,
occupying web page servers with request for irrelevant web pages,
and clogging data network paths with transmissions of irrelevant
web page information.
[0006] As the size of the Internet continues to increase, it
becomes increasingly more desirable to have innovative techniques
for efficiently searching hyperlinked documents.
SUMMARY OF THE INVENTION
[0007] The present invention is directed to a computer-implemented
method for spell-checking text utilizing heuristics. The present
invention is exemplified in a number of implementations and
applications, some of which are summarized below.
[0008] According to an example, embodiment of the present
invention, a computer-implemented application includes a method for
spelling error detection in a target text-string, such as a word or
phrase. The target text-string is compared to a database, or
corpus, to determine a set of contexts which each include an
occurrences of the target text-string. Each context of the set of
contexts is further characterized either including a correct
spelling of the target text-string, including an incorrect spelling
of a reference text-string (e.g., another word or phrase), or being
a context including an indeterminate usage of the target
text-string in the context. A likelihood that the target
text-string is a misspelling of the reference text-string is
thereafter computed as a function of the quantity of contexts
including a correct spelling of the target text-string and the
quantity of contexts including an incorrect spelling of a reference
text-string. In one more particular example implementation of the
present invention, a probability that the target text-string is
misspelled is computed as a ratio of quantity of contexts including
a correct spelling of the target text-string relative to the
quantity of non-indeterminate contexts.
[0009] According to another general example embodiment of the
present invention, a computer-implemented application detects
spelling errors in a target text-string, such as a word or
phrase.
[0010] The target text-string is compared to a database of contexts
to determine from the comparison, a set of
potentially-corresponding contexts; each context in the set having
an "occurrence of the target text-string" characterized as either
including a correct spelling of the target text-string, an
incorrect spelling of a reference text-string, or being an
indeterminate context. Using a quantification of each
characterization, according to the present invention, the computer
application computes a likelihood that the target text-string is
misspelled. For example, with X being the quantity of contexts
including a correct spelling of the target text-string, Y being the
quantity of contexts including an incorrect spelling of a reference
text-string, and Z being the quantity of indeterminate contexts, a
likelihood that the target text-string is a misspelling of the
reference text-string is computed as a function of one of X and Y,
relative to X plus Y. In more typical implementations of the
present invention, each of X, Y and Z is a positive integer. In
another implementation, the computation of likelihood does not
include Z.
[0011] According to other aspects of the present invention,
heuristics are applied to characterize contexts, the heuristics
being a function of the occurrences of the target text-string and
the reference text-string in the context. Contexts are
characterized as including an incorrect spelling of the reference
text-string if occurrences of the reference text-string in the
context are equal to or greater than a pre-determined minimum
quantity threshold (e.g., 1), and a ratio of reference text-string
occurrences in the context to target text-string occurrences in the
context is also equal to or greater than a pre-determined ratio
threshold. Contexts are characterized as including a correct
spelling of the target text-string if occurrences of the target
text-string in the context are equal to or greater than a second
pre-determined quantity threshold (e.g., 1), and a ratio of target
text-string occurrences in the context to reference text-string
occurrences in the context is also equal to or greater than a
second pre-determined ratio threshold. Contexts not characterizable
as either correctly spelled or misspellings are indeterminate.
[0012] According to a further example embodiment of the present
invention, a computer-implemented search engine application detects
spelling errors in a target text-string included within a received
search query.
[0013] In another example embodiment of the present invention, a
method is provided for detecting spelling errors in a target
text-string by selecting a reference text-string having
characteristics corresponding to the target text-string, computing
a first ratio of occurrences of the reference text-string relative
occurrences of the target text-string in a first database,
computing a second ratio of occurrences of the reference
text-string relative to occurrences of the target text-string in a
second database, and determining a likelihood that the target
text-string is misspelled as a function of the first ratio relative
to the second ratio. The first and second databases are each a
corpus including naturally occurring text that are similar in
patterns of content to each other and the text being examined.
However, the second database includes fewer spelling errors than
the first database.
[0014] According to another example embodiment of the present
invention, a target text-string is compared to a database, or
corpus, to determine a set of contexts, each of which includes an
occurrence of the target text-string. Each context of the set of
contexts is further characterized, by first using the corpus and
then using a better-spelled corpus, as either including a correct
spelling of the target text-string, including an incorrect spelling
of a reference text-string (e.g., another word or phrase), or being
a context including an indeterminate usage of the target
text-string in the context. A computation is made of a first ratio
of occurrences of the reference text-string relative occurrences of
the target text-string in the first database. A second ratio is
computed of occurrences of the reference text-string relative to
occurrences of the target text-string in the second database. Using
this computation, the embodiment provides a likelihood that the
target text-string is misspelled is determined as a function of the
first ratio and the second ratio. According to a further aspect,
the target text-string is received as a portion of a search query
for a computer-implemented data network search engine.
[0015] According to another example embodiment of the present
invention, web page information is controlled in response to a user
query identifying a first target web page. Each text-string of the
user query is spell-checked. The resulting correctly-spelled search
query identifies a second target web page. A database is searched
to determine whether the second target web page corresponds to at
least one destination web page. In response to the second target
web page corresponding to said at least one destination web page,
link information is presented for the user to access the
destination web pages, along with peripheral information relevant
for evaluating the link.
[0016] The above summary of the present invention is not intended
to describe each illustrated embodiment or every implementation of
the present invention. The figures and detailed description that
follow more particularly exemplify these embodiments.
BRIEF DESCRIPTION OF THE DRAWINGS
[0017] The invention may be more completely understood in
consideration of the following detailed description of various
embodiments of the invention in connection with the accompanying
drawings, in which:
[0018] FIG. 1 illustrates a system block diagram of an example
embodiment of a data network arrangement, according to the present
invention.
[0019] FIG. 2 illustrates an example embodiment of a corpus
database, according to an example embodiment of the present
invention
[0020] While the invention is amenable to various modifications and
alternative forms, specifics thereof have been shown by way of
example in the drawings and will be described in detail. It should
be understood, however, that the intention is not to limit the
invention to the particular embodiments described. On the contrary,
the intention is to cover all modifications, equivalents, and
alternatives falling within the spirit and scope of the invention
as defined by the appended claims.
DETAILED DESCRIPTION
[0021] The present invention is believed to be applicable to
various types of text pattern recognition methods, including
computer-implemented spell-checking applications within word
processing, speech recognition/transcription, and text-manipulation
programs. The present invention has been found particularly suited
to computer-implemented information searching and retrieval
applications, such as data network search engine applications, for
example. While the present invention is not necessarily limited to
such search engine spell-checking applications, various aspects of
the invention may be appreciated through a discussion of various
examples using this context.
[0022] A great deal of digital information is communicated to
humans via written text displayed on a monitor coupled to a digital
processor. Computer-implemented spell checking routines are
therefore increasingly more desirable as a means to identify
potential text spelling errors. Words are the building blocks of
written and spoken language by which ideas are conveyed. Letters of
the alphabet comprise a pre-defined set of characters used as
phonetic symbols. Particular arrangements of letters are recognized
as words by some accepted authority, one or more meanings (e.g.,
the idea) being associated with the arrangement of letters.
Particular letters, along with their particular order within a
string of text, are important features by which words of different
meaning are distinguished from one another. Typically, the
authority tabulates recognized words and their associated
meaning(s), for example a dictionary publisher.
[0023] Like-sounding words can have different spellings, and
different meanings. In the English language, a context within which
a particular word is used can influence the pronunciation and/or
meaning of the word. Language includes a set of rules supporting
common understanding of word usage, for example grammar. The rules
may be faunal, or very informal, slang being an example of the
later for example. In the English language, words are delineated in
written text by spaces, and in spoken text by pauses between words.
Context of a particular word (e.g., a target word) is the word(s)
adjacent to, or nearby, the target word.
[0024] Spelling errors in text-strings occur due to a number of
reasons. A string of letters (or text-string), delineated as a
word, which is not included in a tabulation of recognized words
(e.g., "silber" instead of "silver"), is a candidate for being a
misspelling of another word. However, the unrecognized text string
may constitute a new word not yet included in the tabulation of
recognized words as having an associated meaning, or may be a
proper name identifying of a particular individual, place or thing.
Often an unrecognized text string is very similar to a recognized
word, for example a word having one or more additional letters,
omitted letters, transposed letters, or substituted letters. A
typing error, where a user depresses an incorrect key, is one
example of a spelling error due to a substituted letter appearing
in a word.
[0025] The English language includes phonetically-similar words
having similar pronunciations, but different spellings (e.g., blue
and blew). Occasionally, a phonetically-identical word is used
intentionally, but incorrectly, in a particular context in place of
a word having the intended meaning. Incorrect usage of
correctly-spelled words is considered another form of spelling
error. The "misspelled" word exists in a tabulation of recognized
words, and therefore is only detectable from the context of the
word's usage, the meaning of the "misspelled" word being
inconsistent with the meaning conveyed by the words around the
"misspelled" word. With the growth of digitally-formatted
information due to an increased usage of computer-implemented
processes, ideas are increasingly transcribed into written form
(e.g., words) as an intermediate step before digitization for
communication and/or storage.
[0026] Text-string spelling errors are one factor contributing to a
poor quality information search query, and consequently, to poor
quality search query results. Spelling errors include misspelled
words and the incorrect usage of correctly spelled words. For
example, a user wishing to search a computer-implemented data
network, such as the Internet, for little red wagons ideally
executes a search on the query "little red wagons" via a search
engine application. However, the user may erroneously enter the
query "little rwd wagons" into a search engine. Obviously, "rwd" is
a misspelling of the intended word "red." The spelling error is
most likely attributable to a typographical error due to the
proximity of the "w" key to the "e" key on a computer keypad used
to input the query. The misspelled word is not recognizable as a
word, and a conventional search engine will typically produce
results including an attempt to locate web pages relevant to the
query text string "rwd." The search is further misdirected since
the search engine does not attempt to locate web pages relevant to
the intended word, "red."
[0027] Search query spelling errors attributable to incorrect usage
of correctly spelled words are illustrated for example by a search
query of "little bed wagons," or even the phonetically-correct
"little read wagons." Each of the afore-mentioned queries include
correctly-spelled words that are incorrectly used. It follows that
the word located between "little" and "wagons" is not spelled
correctly to achieve a correct meaning of the group of words. The
error in the search query is undetectable from a simple
determination of whether each query text-string is a word, for
example, recognized as being a word by being included in a list or
look-up table of recognized words. The errors are detectable by
considering each text-string (e.g., word or phrase) in a context
within which the text-string is used words relative to established
rules for using words within the language.
[0028] In one general example embodiment of the present invention,
a computer-implemented application detects spelling errors in a
target text-string, such as a word or phrase. The target
text-string is compared to a database, or corpus, to determine a
set of contexts which each include an occurrences of the target
text-string. Each context of the set of contexts is further
characterized either including a correct spelling of the target
text-string, including an incorrect spelling of a reference
text-string (e.g., another word or phrase), or being a context
including an indeterminate usage of the target text-string in the
context. A likelihood that the target text-string is a misspelling
of the reference text-string thereafter computed as a function of
the quantity of contexts including a correct spelling of the target
text-string and the quantity of contexts including an incorrect
spelling of a reference text-string. In one more particular example
implementation of the present invention, a probability that the
target text-string is misspelled is computed as a ratio of quantity
of contexts including a correct spelling of the target text-string
relative to the quantity of non-indeterminate contexts.
[0029] According to another general example embodiment of the
present invention, a computer-implemented application detects
spelling errors in a target text-string, such as a word or phrase.
The target text-string is compared to a database of contexts to
determine from the comparison, a set of contexts having an
occurrence of the target text-string. By counting the quantity of
each characterization, according to the present invention, the
computer application computes a likelihood that the target
text-string is misspelled. For example, with X being the quantity
of contexts including a correct spelling of the target text-string,
Y being the quantity of contexts including an incorrect spelling of
a reference text-string, and Z being the quantity of indeterminate
contexts, a likelihood that the target text-string is a misspelling
of the reference text-string is computed as a function of X,
relative to X plus Y. According to other more specific
implementations of the present invention, the computation is a
function of X, relative to X plus Y where each of X, Y and Z is a
positive integer.
[0030] According to another important aspect of the present
invention, heuristics are applied to characterize contexts, the
heuristics being a function of the occurrences of the target
text-string and the reference text-string in the context. Contexts
are characterized as including an incorrect spelling of the
reference text-string if occurrences of the reference text-string
in the context are equal to or greater than a pre-determined
minimum quantity threshold (e.g., 1), and a ratio of reference
text-string occurrences in the context to target text-string
occurrences in the context is also equal to or greater than a
pre-determined ratio threshold. Contexts are characterized as
including a correct spelling of the target text-string if
occurrences of the target text-string in the context are equal to
or greater than a second pre-determined quantity threshold (e.g.,
1), and a ratio of target text-string occurrences in the context to
reference text-string occurrences in the context is also equal to
or greater than a second pre-determined ratio threshold. Contexts
not characterizable as either correctly spelled or misspellings are
classified or tagged "indeterminate."
[0031] According to another general example embodiment of the
present invention, a computer-implemented search engine application
detects spelling errors in a target text-string included within a
received search query.
[0032] According to another general example embodiment of the
present invention, spelling errors are detected in a target
text-string by selecting a reference text-string having
characteristics corresponding to the target text-string, computing
a first ratio of occurrences of the reference text-string relative
occurrences of the target text-string in a first database,
computing a second ratio of occurrences of the reference
text-string relative to occurrences of the target text-string in a
second database, and determining a likelihood that the target
text-string is misspelled as a function of the first ratio relative
to the second ratio. The first and second databases are each a
corpus including naturally occurring text that are similar in
patterns of content to each other and the text being examined.
However, the second database includes fewer spelling errors than
the first database.
[0033] According to another general example embodiment of the
present invention, a target text-string is compared to a database,
or corpus, to determine a set of contexts which each include an
occurrences of the target text-string. Each context of the set of
contexts is further characterized, first using the corpus and
second using a better-spelled corpus, as either including a correct
spelling of the target text-string, including an incorrect spelling
of a reference text-string (e.g., another word or phrase), or being
a context including an indeterminate usage of the target
text-string in the context. A first ratio is computed of
occurrences of the reference text-string relative occurrences of
the target text-string in the first database. A second ratio is
computed of occurrences of the reference text-string relative to
occurrences of the target text-string in the second database, and a
likelihood that the target text-string is misspelled is determined
as a function of the first ratio and the second ratio. According to
a further aspect, the target text-string is received as a portion
of a search query for a computer-implemented data network search
engine.
[0034] According to another general example embodiment of the
present invention, web page information is controlled in response
to a user query identifying a first target web page. Each
text-string of the user query is spell-checked. The resulting
correctly-spelled search query identifies a second target web page.
A database is searched to determine whether the second target web
page corresponds to at least one destination web page. In response
to the second target web page corresponding to said at least one
destination web page, link information is presented for the user to
access the destination web pages, along with peripheral information
relevant for evaluating the link,
[0035] In another example embodiment, the present invention is
direct to a process for estimating a probability that a random
instance of a given text-string (e.g., word or phrase), "bad_word,"
is a misspelling of a reference text-string. The reference
text-string, "good_word," is another phrase, word, or portion
thereof. This probability is expressed in shorthand notation as:
p.sub.Misspell(bad_word, good_word). The bad_word is one
text-string extracted from a text being examined for spelling
errors. A large corpus of naturally occurring text is similar in
patterns of content and misspelling to the text being examined. The
method of the present invention does not require manual tagging or
intervention of the corpus. According to one aspect of the present
invention, occurrences of bad_word in the corpus are broken up into
a set of contexts. Contexts include at least one occurrence of
bad_word and are defined based upon the words located adjacently or
nearby the occurrence of bad_word. For each of these contexts,
heuristics are applied to determine whether the collective
occurrences of bad_word in the context include misspellings of
good_word in the context, include correct spellings of bad_word in
the context, or whether not enough information is available to
distinguish between the correctness and incorrectness of bad_word
(i.e., the context is indeterminate) in the context.
[0036] The probability, p.sub.Misspell(bad_word, good_word), is
estimated as the ratio of the number of instances of bad_word in
contexts (of the set of contexts) characterized as including
misspellings of good_word, to the total number of instances of
bad_word in contexts (of the set of contexts) that were not
characterized to be indeterminate (e.g., misspellings of good_word,
or correct spellings of bad_word in the context). Alternatively,
probability, p.sub.Misspell(bad_word, good_word), is equivalently
estimated as one minus the ratio of the number of instances of
bad_word in contexts characterized as including correct spellings
of bad_word, to the total number of instances of bad_word in
contexts of the set of contexts that were not characterized to be
indeterminate.
[0037] Occurrences of bad_word are determined to be misspellings of
good_word in a given context by comparing of the number of
occurrences of bad_word in the given context, f.sub.bad, to the
number of occurrences of good_word in the given context,
f.sub.good. If f.sub.bad is significant, and greater than
f.sub.good, the occurrences of bad_word in the given context are
deemed to be correctly spelled occurrences of bad_word, and the
given context is determined to include correct spellings of
bad_word. If f.sub.good is significant, and greater than f.sub.bad,
the occurrences of bad_word in the given context are deemed to be
misspellings, and the given context is determined to include
misspellings of good_word. Those contexts including at least one
occurrence of bad_word but not meeting the criteria for
characterizing the context as including either correct spellings of
bad_word or misspellings of good_word are characterized as being
indeterminate. A variety of different heuristics, in addition to
those detailed above, are contemplated by the method of the present
invention.
[0038] According to one example implementation of the present
invention, the number of occurrences of a word in a context is
significant if it is at least meets a given threshold and the ratio
of the frequency of the word in this context to the frequency of
the word in the whole corpus is at least a second threshold.
Thresholds of 3 and 30 have been found to be useful.
[0039] Other significance-determination implementations are
contemplated within the scope of the present invention. In a
further example implementation of significance determinations,
comparison thresholds are determined dynamically according to
pre-determined criteria. According to another example embodiment of
the present invention, a two corpus method of the present invention
includes an dimension added to the method set forth above. The
ratio of the frequencies (of occurrences) of good_word relative to
bad_word (i.e., f.sub.good/f.sub.bad) in a given context are first
determined from a corpus (e.g., a main corpus) as described above,
which ratio shall hereafter be referred to as the "main corpus
ratio". Another ratio of the frequencies (of occurrences) of
good_word relative to bad_word (i.e., f.sub.good/f.sub.bad) in the
given context are also determined from a second, better-spelled
corpus, which ratio shall hereafter be referred to as the
"better-spelled ratio". The better-spelled corpus is similar in
patterns of content to the main corpus, but includes fewer words
are misspelled in the second corpus than are misspelled in the main
corpus. Contexts are determined as described above from either one
of the corpora. Finally, the ratio of (1) the better-spelled ratio
to (2) the main corpus ratio shall hereafter be referred to as the
"better-to-main ratio".
[0040] The main corpus ratio in the given context , the
better-spelled ratio, and the better-to-main ratio are then used to
determine whether bad_word is likely misspelled. For example, if
the main corpus ratio is greater than a given threshold (a
threshold of one works well), if better-spelled ratio is greater
than a given threshold (a threshold of two works well), and if the
better-to-main ratio is greater than a given threshold (a threshold
of two works well), bad_word is likely misspelled in the given
context. If the better-spelled ratio is less than a given threshold
(a threshold of one works well), bad_word is likely spelled
correctly in the given context. Other thresholds, of course, may be
used. Similarly, other comparisons of the main corpus ratio, the
better-spelled ratio, and/or the better-to-main ratio may be
used.
[0041] According to another example embodiment of the present
invention, a two corpus method is applied without restriction to a
context. For example, if the main corpus ratio is greater than a
given threshold (such as one), if the better-spelled ratio is
greater than a given threshold (such as two), and if the
better-to-main ratio is greater than a given threshold (such as
two), bad_word is likely misspelled in this context. If the
better-spelled ratio is less than a given threshold (such as one)
or if the better-to-main ratio is greater than a given threshold
(such as 1.5), then bad world is likely correct in the given
context. Other thresholds, of course, may be used. Similarly, other
comparisons of the main corpus ratio, the better-spelled ratio,
and/or the better-to-main ratio may be used.
[0042] Related words are sometimes misidentified as misspellings in
context-based determinations. According to another example
embodiment of the present invention, these related word
misidentifications are mitigated. If bad_word is truly a universal
(i.e., not context sensitive) misspelling of good_word, bad_word is
expected to occur in every context in which good_word frequently
occurs. Discovery of at least one context in which good_word
appears often, but bad_word occurs very seldom, indicates that
bad_word may not really a universal misspelling of good_word
applicable anywhere. For example, a context-based determination may
conclude that "woman" is a misspelling of "women," based on
frequencies of occurrence heuristics, such as those set forth
above, in a significant portion of the same contexts. However,
observing that the phrase "What Women Want" (a popular movie title)
occurs frequently, but that "What Woman Want" almost never occurs,
indicates that in certain contexts, "woman" is not a misspelling of
"women." Therefore, "woman" is not a universal misspelling of
"women" and the present example method concludes that the target
text-string (i.e., "woman") is never really a misspelling of the
reference text-string (i.e., "women") in any context. In an
alternative implementation, discovery of at least N contexts in
which good_word appears often, but bad_word occurs very seldom, is
necessary to indicate that bad_word may not really a universal
misspelling of good_word applicable anywhere, N being greater than
one.
[0043] According to a further example embodiment of the present
invention, the above-described two corpus, context insensitive
method is used in combination with one of the context sensitive
methodologies described previously, to supervise the determination
of misspelling likelihood. For example, a likelihood that a target
text-string is misspelled is computed unless at least N contexts
are discovered in which the reference text-string appears often,
but a target text-string occurs very seldom, indicating that the
target text-string is not really a universal misspelling of the
reference text-string.
[0044] Use of contexts in the example embodiments of the present
invention for determining the likelihood (i.e.,
p.sub.Misspell(bad_word, good_word)) that a target word is
misspelled is distinguishable from the general use of n-gram
contextual information in conventional statistical language model
(SLM) based spelling correction methods. For example, other
SLM-based spelling correction methods might correctly correct
"collage cheerleaders" to "college cheerleaders" based on a
frequency of "college cheerleaders" in a training corpus. However,
the other SLM-based spelling correction methods typically may not
be able to correct "collage" to "college" in isolation, or in a
novel context (i.e., not a well-known context such as "college
cheerleaders"). The method of the present invention is capable of
determining that "collage" is usually a misspelling of "college" by
determining a likelihood of a "universal" misspelling derived from
a variety of contexts, and ensuring the likelihood determination is
accurate through application of significance determinations.
[0045] According to a further example embodiment of the present
invention, the above-described example embodiments are used to
compute a likelihood (e.g., probability) that a target text-string
is a misspelling of a reference text-string. The probability is
thereafter used in a reverse manner to select or suggest the
reference text-string as a potential spelling correction for the
target text-string to the user. In one example implementation, the
user is presented with a ranked list of alternative reference
text-strings from which a selection may be made for substitution
with the target text-string being examined.
[0046] According to a further example embodiment of the present
invention, the above-described methods are implemented in a data
network search engine application to check the spelling of text
used to direct a search of the data network. For example, an
Internet search engine includes spell-checking steps, such as those
outlined above. The search engine prompts a user for a search
query, the query being a series of text-strings identifying a first
target web page. The text-strings of the search query are subjected
to a spell-check examination. In one implementation, the user is
prompted to confirm or correct text-strings identified as likely
misspelled. In another implementation, spelling errors are
automatically corrected using a reference text-string having a
significantly large likelihood of being a misspelling. The search
query resulting following the spelling-checking procedures
identifies a second target web page. The search engine subsequently
conducts a search of the data network for at least one destination
web page based upon the spell-corrected search query identifying
the second target web page.
[0047] FIG. 1 illustrates an one example of a computer system 100
implementing a spell-checking function of the present invention. A
user's computer 110 is coupled to a data network (e.g., the
Internet) 180 via a communication interface 115. Web page servers,
120, 130 and 140 respectively are also coupled to data network 180,
each being adapted to serve web pages (e.g., hyperlinked documents)
and other information through the data network. For example, server
120 hosts web pages 125, server 130 hosts web pages 135, and server
140 hosts web pages 145. A browser application executing on user
computer 110 facilitates retrieval of data network information, web
pages for example.
[0048] Computing apparatus 150 is coupled to the data network.
Computer 150 includes a storage media 160 coupled to a processor
170. Storage media 160 stores at least one database 165, the
database being a spell-checking corpus, for example. In one example
arrangement, computer 150 executes, through processor 170, a data
network search engine application. The search engine application is
adapted to search the data network for destination web pages in
response to a search query. Searching the data network includes
searching a web page summary database stored in storage media, the
database including descriptive and linking information
characterizing each of web pages 120, 130, and 140
respectively.
[0049] A user submits a search query through user computer 110, the
communication interface 115 and data network 180 to the search
engine application running on computer 150. The search query
identifies a first target web page. First target web page
corresponds to web page 120, for example. However, the search query
may contain spelling errors in the text of the search query that
misidentifies a user-intended destination web page, for example the
user may be attempting to retrieve web page 140 through their
search. The search engine application includes a spell-checking
method, for example as described herein, which is used to
spell-check text of the received search query. The
post-spell-checked search query identifies a second target web
page, for example, second target web page corresponds to
user-intended destination web page 140. It is possible for second
target web page to be different than first target web page if
significant changes are made to text terms of the search query as a
result of the spelling error detection process. Alternatively,
second target web page may not be different than first target web
page if no, or insignificant, changes are made to the search query
as a result of the spelling error detection process. The search
engine executes a search of the data network, and/or the summary
database, responsive to the spell-checked search query to determine
whether the identified second target web page corresponds to at
least one destination web page, web page 140 for example.
Thereafter, the search engine presents link and descriptive web
page summary information to the user based upon search results.
[0050] FIG. 2 illustrates one example embodiment of a corpus of the
present invention. A database (e.g., a corpus) 210 is stored in a
storage media 200. FIG. 2 shows excerpts of text included in the
corpus. From the text, various contexts are determinable based upon
a target text-string (e.g., a word). One example of a target
text-string 220 is the word "red." The corpus may include one or
more occurrences of target text-string 220 as is shown in FIG. 2.
Contexts of the target text-string are determined from occurrences
of the target text-string in the corpus. For example, a first
context 230 is derived from the target word following the word
"little." The context is a text-string following the word "little."
Other examples of the context are illustrated in FIG. 2 at 230'
(another occurrence of the target word "red" following the word
"little"), 230'' (the same context although the word "little" is
followed by a different word, "bed"), and 230''' (the word "little"
followed by a word, "blue"). Alternatively, the context is
definable as a text-string being within 1 word of the word
"little," which would also include the word arrangement indicated
at 232 where a word 250, "read," precedes the word "little," in
addition to the context examples indicated at 230, 230', 230'', and
230''' previously discussed. Also derived from an occurrence of the
target word "red" in the corpus, a second context 225 defined as a
word preceding the word "paint." A third context 240 is also
definable, context 240 being the target word situated between the
words, "little" and "wagons." While no other examples of the second
context 225 are illustrated in FIG. 2, two other examples of the
third context 240 are illustrated, including 240' (another
occurrence of the target word "red" situated between "little" and
"wagons") and 240'' (a different word 252, "blue," but situated
between "little" and "wagons"). Words "read" 250, and "blue" 252
are illustrations of reference text-strings (e.g., words).
Reference text-strings are other text-strings that occur in
contexts identified from occurrences of the target text-string.
According to the method of the present invention, heuristics are
used to determine whether the target text-string is a misspelling
of reference text-strings identified from contexts of the target
text-string.
[0051] Accordingly, the present invention is not to be necessarily
limited to the particular examples described above, but is intended
to cover all aspects of the invention as fairly set out in the
attached claims. For instance, while a method for spell-checking
words of a search query of a data network search engine application
is illustrated, other techniques for spell-checking text-strings in
computer-implemented applications can benefit from the above
mentioned teachings. Various modifications, equivalent processes,
as well as numerous structures to which the present invention may
be applicable will be readily apparent to those of skill in the art
to which the present invention is directed upon review of the
present specification. The claims are intended to cover such
modifications and devices.
* * * * *