Spelling Correction for Search Engine Queries.
ABSTRACT Search engines have become the primary means of accessing informa- tion on the Web. However, recent studies show misspelled words are very com- mon in queries to these systems. When users misspell query, the results are incor- rect or provide inconclusive information. In this work, we discuss the integration of a spelling correction component into tumba!, our community Web search en- gine. We present an algorithm that attempts to select the best choice among all possible corrections for a misspelled term, and discuss its implementation based on a ternary search tree data structure.
- [Show abstract] [Hide abstract]
ABSTRACT: In this paper we present a language-independent approach for conflation that does not depend on predefined rules or prior knowledge of the target language. The proposed unsupervised method is based on an enhancement of the pure n-gram model that can group related words based on various string-similarity measures, while restricting the search to specific locations of the target word by taking into account the order of n-grams. We show that the method is effective to achieve high score similarities for all word-form variations and reduces the ambiguity, i.e., obtains a higher precision and recall, compared to pure n-gram-based approaches for English, Portuguese, and Arabic. The proposed method is especially suited for conflation approaches in Arabic, since Arabic is a highly inflectional language. Therefore, we present in addition an adaptive user interface for Arabic text retrieval called “araSearch”. araSearch serves as a metasearch interface to existing search engines. The system is able to extend a query using the proposed conflation approach such that additional results for relevant subwords can be found automatically. © 2009 Wiley Periodicals, Inc.Journal of the American Society for Information Science and Technology 01/2009; 60(7):1448-1465. · 2.01 Impact Factor
- [Show abstract] [Hide abstract]
ABSTRACT: Figures are important experimental results that are typically reported in full-text bioscience articles. Bioscience researchers need to access figures to validate research facts and to formulate or to test novel research hypotheses. On the other hand, the sheer volume of bioscience literature has made it difficult to access figures. Therefore, we are developing an intelligent figure search engine (http://figuresearch.askhermes.org). Existing research in figure search treats each figure equally, but we introduce a novel concept of "figure ranking": figures appearing in a full-text biomedical article can be ranked by their contribution to the knowledge discovery. We empirically validated the hypothesis of figure ranking with over 100 bioscience researchers, and then developed unsupervised natural language processing (NLP) approaches to automatically rank figures. Evaluating on a collection of 202 full-text articles in which authors have ranked the figures based on importance, our best system achieved a weighted error rate of 0.2, which is significantly better than several other baseline systems we explored. We further explored a user interfacing application in which we built novel user interfaces (UIs) incorporating figure ranking, allowing bioscience researchers to efficiently access important figures. Our evaluation results show that 92% of the bioscience researchers prefer as the top two choices the user interfaces in which the most important figures are enlarged. With our automatic figure ranking NLP system, bioscience researchers preferred the UIs in which the most important figures were predicted by our NLP system than the UIs in which the most important figures were randomly assigned. In addition, our results show that there was no statistical difference in bioscience researchers' preference in the UIs generated by automatic figure ranking and UIs by human ranking annotation. The evaluation results conclude that automatic figure ranking and user interfacing as we reported in this study can be fully implemented in online publishing. The novel user interface integrated with the automatic figure ranking system provides a more efficient and robust way to access scientific information in the biomedical domain, which will further enhance our existing figure search engine to better facilitate accessing figures of interest for bioscientists.PLoS ONE 01/2010; 5(10):e12983. · 3.53 Impact Factor
- [Show abstract] [Hide abstract]
ABSTRACT: Interaction faults caused by a flawed external system designed by a third party are a major issue faced by interconnected systems. In this paper, we define a scenario where this type of problem occurs, and describe some fault cases observed in real systems. We also discuss the most important challenges faced in this scenario, focusing on error detection. The problem is divided in several sub-problems, some of which can be addressed by traditional or simple techniques, and some of which are complex problems by themselves. The purpose of this paper is not to present ad hoc solutions to specific sub-problems, but to introduce a new scenario and give general approaches to address each sub-problem. That includes a detailed insight on important concepts, such as implicit redundancies. With this, we lay down the foundations for a wide range of future work.Service Availability, 5th International Service Availability Symposium, ISAS 2008, Tokyo, Japan, May 19-21, 2008, Proceedings; 01/2008
Spelling Correction for Search Engine Queries
Bruno Martins and Mário J. Silva
Departamento de Informática
Faculdade de Ciências da Universidade de Lisboa
1749-016 Lisboa, Portugal
Abstract Search engines have become the primary means of accessing informa-
tion on the Web. However, recent studies show misspelled words are very com-
mon in queries to these systems. When users misspell query, the results are incor-
rect or provide inconclusive information. In this work, we discuss the integration
of a spelling correction component into tumba!, our community Web search en-
gine. We present an algorithm that attempts to select the best choice among all
possible corrections for a misspelled term, and discuss its implementation based
on a ternary search tree data structure.
Millions of people use the Web to obtain needed information, with search engines cur-
rently answering tens of millions of queries every day. However, with the increasing
popularity of these tools, spelling errors are also increasingly more frequent. Between
10 to 12 percent of all queryterms entered into Web search engines are misspelled .
A large number of Web pages also contain misspelled words. Web search is thus a task
of information retrieval in an environment of faulty texts and queries. Even with mis-
spelledterms in the queries,searchenginesoftenretrieveseveralmatchingdocuments–
those containingspelling errors themselves. However,the best and most “authoritative”
pages are often missed, as they are likely to contain only the correctly spelled forms.
An interactive spelling facility that informs users of possible misspells and presents ap-
propriate corrections to their queries could bring improvements in terms of precision,
recall, and user effort. Google was the first major search engine to offer this facility .
Oneof thekeyrequirementsimposedbythe Web environmentona spelling checker
is that it should be capable of selecting the best choice among all possible corrections
for a misspelled word, instead of giving a list of choices as in word processor spelling
checking tools. Users of Web search systems already give little attention to query for-
mulation, and we feel that overloading them with an interactive correction mechanism
would not be well accepted. It is therefore importantto make the right choice among all
possible corrections autonomously.
This work presents the developmentof a spelling correction component for tumba!,
our community search engine for the Portuguese Web . In tumba! we check the
query for misspelled terms while results are being retrieved. If errors are detected, we
provide a suggestive link to a new “possibly correct” query, together with the search
results for the original one.
The rest of this paper is organized as follows: the next section presents the termi-
nology used throughoutthis work. Section 3 gives an overview on previous approaches
to spelling correction. Section 4 presents the ternary search tree data structure, used in
our system for storing the dictionary. Section 5 details our algorithm and the heuristics
behind it. Section 6 describes the data sources used to build the dictionary. Section 7
describes experimental results. Finally, Section 8 points our conclusions and directions
for future work.
Information Retrieval (IR) concerns with the problem of providing relevant docu-
ments in response to a user’s query . The most commonly used IR tools are Web
search engines, which have become a fact of life for most Internet users. Search en-
ginesuse softwarerobotstosurveytheWeb, retrievingandindexingHTMLdocuments.
Queries are checked against the keyword indexes, and the best matches are returned.
Precision and Recall are the most popular metrics in evaluating IR systems. Pre-
cision is the percentage of retrieved documents that the searcher is actually interested
on. Recall, on the other hand, is the percentage of relevant documents retrieved from
the set of all documents, this way referring to how much information is retrieved by the
search. The ultimate goal of an information retrieval system is to achieve recall with
Spelling has always been an issue in computer-basedtext tools. Two main problems
can be identified in this context: Error detection, which is the process of finding mis-
spelled words, and Error correction, which is the process of suggesting correct words
to a misspelled one. Although other approaches exist, most spelling checking tools are
based on a dictionary which contains a set of words which are considered to be cor-
rect. The problem of spelling correction can be defined abstractly as follows: Given an
alphabet σ, a dictionary D consisting of strings in σ∗and a string s, where s / ∈ D and
s ∈ σ∗, find the word w ∈ D that is most likely to have been erroneously input as s.
occur because the typist accidentally presses the wrong key, presses two keys, presses
the keys in the wrong order, etc; and phonetic errors, where the misspelling is pro-
nounced the same as the intended word but the spelling is wrong. Phonetic errors are
harder to correct because they distort the word more than a single insertion, deletion or
substitution. In this case, we want to be able to key in something that sounds like the
misspelled word (a “phonetic code”) and perform a “fuzzy” search for close matches.
The search for candidate correct forms can be done at typographic level, and then re-
fined using this method.
Web information retrieval systems have been around for quite some time now, hav-
ing become the primary means of accessing information the Web [1,8]. Early systems
engines did not check query spelling but since April 2001, several webwide search
engines, including Excite and Google, provide dynamic spelling checking, while oth-
ers such as Yahoo, simply tracked common misspellings of frequent queries, such as
movie star names. Technical details for these systems are unavailable, but they seem to
be based on spelling algorithms and statistical frequencies.
Algorithmictechniquesfor detectingandcorrectingspellingerrors in text has also a
long and robust history in computer science . Previous studies have also addressed
the use of spelling correctors in the context of user interfaces . Spelling checkers
(sometimes called “spell checkers” by people who need syntax checkers) are nowadays
common tools for many languages, and many proposals can also be found on the liter-
ature. Proposed methods include edit distance [11,31,21], rule-based techniques ,
n-grams [25,33]probabilistic techniques , neural nets [28,6,17], similarity key tech-
niques [34,23], or combinations [16,22]. All of these methods can be thought of as
calculating a distance between the misspelled word and each word in the dictionary.
The shorter the distance, the higher the dictionary word is ranked as a good correction.
Figure1. The four most common spelling errors.
Edit distance is a simple technique. The distance between two words is the num-
ber of editing operations required to transform one into another. Analysis of errors –
mainly typing errors – in very large text files have found that the great majority of
wrong spellings (80-95%) differ from the correct spellings in just one of the four ways
described in Figure 1. The editing operations to consider should therefore correspond
to these four errors, and candidate corrections include the words that differ from the
original in a minimum number of editing operations . Recent works are experi-
mentingwith modelingmorepowerfuledit operations,allowinggenericstring-to-string
edits . Additional heuristics are also typically used to complement techniques based
on edit distance. For instance, in the case of typographic errors, the keyboard layout is
very important. It is much more usual to accidentally substitute a key by another if they
are placed near each other on the keyboard.
Similarity key methods are based on transforming words into similarity keys that
reflect their characteristics. The words in the dictionary and the words to test are both
transformed into similarity keys. All words in the dictionary sharing the same key with
a word being tested are candidates to return as corrections. An example of this method
is thepopularSoundexsystem.Soundex(thenamestands for“Indexingon sound”)was
devised to help with the problem of phonetic errors [12,19]. It takes an English word
and producesa fourdigit representation,in a rough-and-readyway designedto preserve
the salient features of the phonetic pronunciation of the word.
The metaphone algorithm is also a system for transforming words into codes based
letter scheme, metaphone analyzes both single consonants and groups of letters called
diphthongs, according to a set of rules for grouping consonants, and then mapping
groups to metaphone codes. The disadvantage of this algorithm is that it is specific
to the English language. A version of these rules for the Portuguese language has, to
the best of our knowledge, not yet been proposed. Still, there has been recent research
on machine learning methods for letter-to-phoneme conversion [15,30]. Application of
these techniques to Portuguese should be straightforward, providing he have enough
More recent studies on error correction propose the use of context, attempting to
detect words which are misused but spelled correctly [5,14]. Spelling checkers based
on isolated word methodswould see the followingsentenceas correct: a paragraphcud
half mini flaws but wood bee past by the spill checker. However,since in search engines
users oddly type more than tree terms for a query, it would be a waste to make context
dependent correction. Isolated word methods should prove sufficient for our task.
4Ternary Search Trees
In this work, we use a ternary search tree (TST) data structure for storing the dictio-
nary in memory. TSTs are a type of trie that is limited to three children per node [3,4].
Trie is the common definition for a tree storing strings, in which there is one node for
everycommonprefix and the strings are stored in extraleaf nodes. TSTs have been suc-
cessfully used for several years in searching dictionaries. Search times in this structure
are O(log(n)+k) in the worst case, where n is the number of strings in the tree and k
is the length of the string being searched for. In a detailed analysis of various imple-
mentations of trie structures, the authors concluded that “Ternary Search Tries are an
effective data structure from the information theoretic point of view since a search costs
typically about log(n) comparisons on real life textual data. [...] This justifies using
ternary search tries as a method of choice for managing textual data” .
Figure 2 illustrates a TST. The structure stores key-value pairs, where keys are the
words andvalues are integerscorrespondingto the word frequency.As we cansee, each
node of the tree stores one letter and has three children. A search compares the current
character in the search string with the character at the node. If the search character
comes lexically first, the search goes to the left child; if the search character comes
after, the search goes to the right child. When the search character is equal, the search
goes to the middle child, and proceeds to the next character in the search string.
TSTs combine the time efficiency of tries with the space efficiency of binary search
trees. They are faster than hashing for many typical search problems, and support a
broad range of useful operations, like finding all keys having a given prefix, suffix, or
infix, or finding those keys that closely match a given pattern.
Figure2. A ternary search tree storing the words “to”, “too”, “toot”, “tab” and “so”, all with an
associated frequency of 1.
5Spelling Correction Algorithm
A TST data structure stores the dictionary. For each stored word, we also keep a fre-
possible corrections for a misspelled word, we use these word frequency counts as a
popularity ranking, together with other information such as metaphone keys. Although
we donot have a specific text-to-phonemealgorithmfor the Portugueselanguage,using
the standard metaphone algorithm yields in practice good results.
Queries entered in the search engine are parsed and the individual terms are ex-
tracted, with non word tokens ignored. Each word is then converted to lower case, and
checked to see if it is correctly spelled. Correctly spelled words found in user queries
are updated in the dictionary, by incrementing their frequency count. This way, we use
the information in the queries as feedback to the system, and the spelling checker can
adapt to the patterns in user’s searches by adjusting its behavior. For the misspelled
words, a correctly spelled form is generated. Finally, a new query is presented to the
user as a suggestion, together with the results page for the original query. By clicking
on the suggestion, the user can reformulate the query.
Our system integrates a wide range of heuristics and the algorithm used for mak-
ing the suggestions for each misspelled word is divided in two phases. In the first, we
generate a set of candidate suggestions. In the second, we select the best.
The first phase of the algorithm can be further decomposed into 9 steps. In each
step, we look up the dictionary for words that relate to the original misspelling, under
1. Differ in one character from the original word.
2. Differ in two characters from the original word.
3. Differ in one letter removed or added.
4. Differ in one letter removed or added, plus one letter different.
5. Differ in repeated characters removed.
6. Correspond to 2 concatenated words (space between words eliminated).
7. Differ in having two consecutive letters exchanged and one character different.
8. Have the original word as a prefix.
9. Differ in repeated characters removed and 1 character different
In each step, we also move on directly to the second phase of the algorithm if one
or more matching words are found (i.e., if there are candidate correct forms that only
differ in one character from the original misspelled word, a correct form that differs in
more characters and is therefore more complex will never be chosen).
In the second phase, we start with a list of possible corrections.We then try to select
the best one, following these heuristics:
1. If there is one solution that differs only in accented characters, we automatically
return it. Typing words without correct accents is a very common mistake in the
Portuguese language (20% according to Medeiros ).
2. If there is one solution that differs only in one character, with the error correspond-
ing to an adjacent letter in the same row of the keyboard (the QWERTY layout is
assumed), we automatically return it.
3. If there are solutions that have the same metaphone key as the original string, we
return the smallest one, that is, the one with less characters.
4. If there is one solution that differs only in one character, with the error correspond-
ing to an adjacent letter in an adjacentrow of the keyboard,we automaticallyreturn
5. In the last case, we return the smallest word.
We follow the list of heuristics sequentially, and only move to the next if no match-
ing words are found. If there is more than one word satisfying the conditions for each
heuristic, we first try to return the one where the first character is equal to the correctly
spelled word. If there is still more than one word, we return the one that has the highest
6Data Sources and Associated Problems
The dictionary for the spelling checking system is a normal text file, where each line
containsa termandits associatedfrequency.Thesources ofPortuguesewordsandword
frequencies for the dictionary were the texts from the Natura-Publico and the Natura-
Minhocorpora[27,26]. The first oneis made of the two first paragraphsof news articles
from Publico in the years of 1991, 1992, 1993 and 1994. The second, corresponds to
the full articles in 86 days of editions of the newspaper Diario do Minho, spread across
the years of 1998 and 1999.
The dictionary strongly affects the quality of a spelling checking system. If it is too
small, not only will the candidate list for misspellings be severely limited, but the user
will also be frustrated by too many false rejections of words that are correct. On the
other hand, a lexicon that is too large may not detect misspellings when they occur, due
to the dense “word space”.
News articles capture the majority of the words commonly used, as well as tech-
nical terms, proper names, common foreign words, or references to entities. However,
such large corpora often contain many spelling errors . We use word frequencies to
choose among possible corrections, which to some extent should deal with this prob-
lem. As misspelled terms are, in principle, less frequent over the corpus than their cor-
responding correct form, only on rare occasions should the spelling checker provide an
The Web environment introduces difficulties. It is general in subject, as opposed to
domain specific, and multilingualism issues are also common. While spelling checkers
in text editors use standard and personal dictionaries, search engine spelling checkers
should be more closely tied to the content they index, providing suggestions based on
the content of the corpus. This would avoid the dead-end effect of suggesting a word
that is correctly spelled but not included in any words on the site, and add access to
names and codes which will not be in any dictionary. However, using a search engine’s
inverted index as the basis of the spelling dictionary only works well when the content
has been copy edited, or when an editor is available to check the word list and reject
Some experiments were performed in order to quantitatively evaluate our spelling cor-
We were first interested in evaluating the quality of the proposed suggestions. To
achieve this, we compared the suggestions produced by our spelling checker against
Aspell – see the project homepage at http://aspell.sourceforge.net/. Aspell is
a popular interactive spelling checking program for Unix environments. Its strength
comes from merging the metaphone algorithm with a near miss strategy, this way cor-
recting phonetic errors and making better suggestions for seriously misspelled words.
The algorithm behind Aspell is therefore quite similar to the one used in our work, and
the quality of the results in both systems should be similar.
das da Lingua Portuguesa (http://ciberduvidas.sapo.pt/\-php/\-glossario.
php) and by inspecting the query logs for the search engine. The table below shows the
list of misspelled terms used, the correctly spelled word, and the suggestions produced.
In the table, a “*” means that the algorithm did not detect the misspelling and a “-”
means the algorithm failed in returning a suggestion.
Correct Form Spelling Error Our Algorithm
Continued on next page
Table 1 – continued from previous page
Correct Form Spelling Error Our Algorithm
despretensioso despretencioso despretensioso despretensioso
indispensável indespensável indispensável indispensável
Continued on next page
Table 1 – continued from previous page
Correct Form Spelling Error Our Algorithm
quadruplicado quadriplicado quadruplicado quadruplicado
miscigenação miscegenação miscigenação miscigenação
Aspell by a slight margin of 1.66%. On the 120 misspellings, our algorithm failed in
detecting a spelling error 38 times, and it failed on providinga suggestion only 5 times.
Note that the data source used to build the dictionaryhas itself spelling errors.A careful
process of reviewing the dictionary could improve results in the future.
Kukich points out that most researchers report accuracy levels above 90% when
the first three candidates are considered instead of the first guess . Guessing the one
rightsuggestionto presentto theuser is muchharderthansimplyidentifyingmisspelled
words and present a list of possible corrections.
In the second experiment, we took some measures from the integration of our
spelling checker with a search engine for the Portuguese Web. We tried to see if, by
using the spelling correction component, there were improvements in terms of preci-
sion and recall in our system. Using a hand compiled list of misspelled queries, we
measured the number of retrieved documents in the original query, and the number of
retrieved documents in the transformed query. We also had an human evaluator access-
ing the quality of the first ten results returned by the search engine, that is, measuring
how many documents in the first ten results were relevant to the query.
Misspelled Query # Relevant Results Correct Query # Relevant Results
Table 2. Results from the Integration of the Spelling Checker With Tumba!
Results confirm our initial hypothesis that integrating spelling correction in Web
search tools can bring substantial improvements. Although many pages were returned
in responsetomisspelled queries(andinsomecases all pageswere indeedrelevant),the
results for the correctly spelled queries were always of better quality and more relevant.
8Conclusions and Future Work
This paper presented the integration of a spelling correction component into tumba!, a
Portuguese community Web search engine. The key challenge in this work was deter-
mining how to pick the most appropriate spelling correction for a mistyped query from
a number of possible candidates.
The spelling checker uses a ternary search tree data structure for storing the dictio-
nary. As source data, we used a large textual corpus of from two popular Portuguese
newspapers. The evaluation showed that our system gives results of acceptable quality,
and that integrating spelling correction in Web search tools can be beneficial. However,
the validation work could be improved with more test data to support our claims.
An important area for future work concerns phonetic error correction. We would
like to experiment with machine learning text-to-phoneme techniques that could adapt
to the Portuguese language, instead of using the standard metaphonealgorithm [15,30].
We also find that queries in our search engine often contain companynames, acronyms,
foreign words and names, etc. Having a dictionary that can account for all these cases
is very hard, and large dictionaries may result in inability to detect misspellings due to
the dense “word space”. However, keeping two separate dictionaries, one in the TST
used for correction and another in an hash-table used only for checking valid words,
could yield interesting results. Studying ways of using the corpus of Web pages and the
logs from our system, as the basis for the spelling checker, is also a strong objective for
future work. Since our system imports dictionaries in the form of ASCII word lists, we
do however have an infrastructure that facilitates lexicon management.
Special thanks to our colleagues and fellow members of the tumba! development, and
to the various members of Linguateca, for their valuable insights and suggestions. This
researchwas partiallysupportedbytheFCCN - Fundaçãoparaa ComputaçãoCientífica
Nacional, FCT - Fundação para a Ciência e Tecnologia, and FFCUL - Fundação da
Faculdadede Ciências da Universidadede Lisboa, undergrants POSI/SRI/47071/2002-
(project GREASE) and SFRH/BD/10757/2002 (FCT scholarship).
1. A. Arasu, J. Cho, H. Garcia-Molina, A. Paepcke, and S. Raghavan. Searching the web. ACM
Transactions on Internet Technology, 1(1):2–43, 2001.
2. R. Baeza-Yates and B. Ribeiro-Neto. Modern Information Retrieval. ACM Press, 1999.
3. J. Bentley and R. Sedgewick. Fast algorithms for sorting and searching strings. In Proceed-
ings of SODA-97, the 8th ACM-SIAM Symposium on Discrete Algorithms, 1997.
4. J. Bentley and R. Sedgewick. Ternary search trees. Dr. Dobb’s Journal, 23(4):20–25, April
5. J. Bigert. Probabilistic detection of context-sensitive spelling errors. In Proceedings of
LREC-2004,the4thInternational Conference onLanguage Resources and Evaluation, 2004.
6. A. G. Bonfante. Uso de redes neurais para correção gramatical do português: Um estudo de
caso. Master’s thesis, Instituto de Ciências Matemáticas e da Computação da Universidade
de São Paulo, São Carlos, São Paulo, Brazil, 1997. Dissertação de Mestrado.
7. E. Brill and R. C. Moore. An improved error model for noisy channel spelling correction.
In Proceedings of ACL-2000, the 38th Annual Meeting of the Association for Computational
Linguistics, pages 286–293, 2000.
8. S.Brinand L.Page. Theanatomy of alarge-scalehypertextual Websearch engine. Computer
Networks and ISDN Systems, 30(1–7):107–117, 1998.
9. J. Clément, P. Flajolet, and B. Vallée. The analysis of hybrid trie structures. In Proceedings
of DA-98, the 9th annual ACM-SIAM symposium on discrete algorithms, pages 531–539.
Society for Industrial and Applied Mathematics, 1998.
10. H. Dalianis. Evaluating a spelling support in a search engine. In Proceedings of NLDB-2002,
the 7th International Workshop on the Applications of Natural Language to Information
Systems, June 2002.
11. F. J. Damerau. A technique for computer detection and correction of spelling errors. Com-
munications of the ACM, 7(3):171–176, 1964.
12. L. Davidson. Retrieval of mis-spelled names in an airline passenger record system. Commu-
nications of the ACM, 5(3):169–171, March 1962.
13. I. Durham, D. A. Lamb, and J. B. Saxe. Spelling correction in user interfaces. Communica-
tions of the ACM, 26(10):764–773, October 1983.
14. M. A. Elmi and M. Evens. Spelling correction using context. In C. Boitet and P. Whitelock,
editors, Proceedings of the 26th Annual Meeting of the Association for Computational Lin-
guistics and 17th International Conference on Computational Linguistics, pages 360–364,
San Francisco, California, 1998. Morgan Kaufmann Publishers.
15. W. M. Fisher. A statistical text-to-phone function using n-grams and rules. In Proceedings
of ICASSP-99, the 1999 IEEE International Conference on Acoustics, Speech and Signal
Processing, volume 2, pages 649–652, March 1999.
16. V. J. Hodge and J. Austin. An evaluation of phonetic spell checkers. Technical Report YCS
338, Department of Computer Science of the University of York, 2001.
17. V. J. Hodge and J. Austin. A novel binary spell checker. In Proceedings of ICANN-01, the
11th International Conference on Articial Neural Networks, August 2001.
18. R. L. Kashyap and J. Oommen. Spelling correction using probabilistic methods. Pattern
Recognition Letters, 1985.
19. D. E. Knuth. The Art of Computer Programming, volume 3 / Sorting and Searching.
Addison-Wesley Publishing Company, Reading, Massachusetts, 2nd edition, 1982.
20. K. Kukich. Techniques for automatically correcting words in text. ACM Computing Surveys,
21. V. I. Levenshtein. Binary codes capable of correcting deletions, insertions and reversals.
Soviet Physics Doklady, 10:707–710, 1966.
22. J. C. D. Medeiros. Processamento morfológico e correcção ortográfica do português. Mas-
ter’s thesis, Instituto Superior Técnico, 1995.
23. L. Philips. Hanging on the metaphone. Computer Language, 7(12):39–43, 1990.
24. L. Philips. The double-metaphone search algorithm. C/C++ User’s Journal, 18(6), June
25. E. M. Riseman and A. R. Hanson. A contextual postprocessing system for error correction
using binary n-grams. IEEE Transactions on Computer Systems, C-23(5):480–493, May
26. D. Santos and P. Rocha. Evaluating cetempúblico, a free resource for portuguese. In Pro-
ceedings of ACL-2001, the 39th Annual Meeting of the Association for Computational Lin-
guistics, pages 442–449, July 2001.
27. D. Santos and L. Sarmento. O projecto AC/DC: acesso a corpora / disponibilização de
corpora. In A. Mendes and T. Freitas, editors, Actas do XVIII Encontro da Associação
Portuguesa de Linguística, pages 705–717, October 2002.
28. T. J. Sejnowski and C. R. Rosenberg. Parallel networks that learn to pronounce english text.
Complex Systems, 1:145–168, 1987.
29. M. J. Silva. The case for a portuguese Web search engine. DI/FCUL TR 03–03, Department
of Informatics, University of Lisbon, March 2003.
30. K. Toutanova and R. C. Moore. Pronunciation modeling for improved spelling correction,
31. R. A. Wagner and M. J. Fischer. The string-to-string correction problem. Communications
of the ACM, 1(21):168–173, 1974.
32. E. J. Yannakoudakis. Expert spelling error analysis and correction. In K. P. Jones, edi-
tor, Proceedings of a Conference held by the Aslib Informatics Group and the Information
Retrieval Group of the British Computer Society, pages 39–52, March 1983.
33. E. M. Zamora, J. J. Pollock, and A. Zamora. The use of trigram analysis for spelling error
detection. Information Processing and Management, 6(17):305–316, 1981.
34. J. Zobel and P. Dart. Phonetic string matching: Lessons from information retrieval. In Pro-
ceedings of SIGIR-96, the 19th Annual International ACM SIGIR Conference on Research
and Development in Information Retrieval, pages 166–172, 1996.