2. substitutions of lower case for upper case at the be­ginning of a proper noun,

a) the percentage of errors in the first letter of a word is higher than that reported in other studies,

b) keyboard adjacency effects are less important than other factors, and

The corpus used contains over 8 million words of ed­ited and unedited texts. It comprises three different sets of texts. The first one, with about 4 million words, is a bal­anced set of edited and unedited texts. The second one is a highly edited corpus with about 2 million words. The third one, with about 2 million words, contains unedited data from the Original Works Creation (OWC) sites. Table 1 below shows the breakdown for each set.

The corpus was analyzed morphologically using a knowledge-based syntactic parser (the parser which un­derlies the Microsoft Office grammar checkers developed by Microsoft). This task produced a list of unknown words. The total number of unique unknown words dis­covered in the corpus was 76K, of which almost 27K (over 35%) were unique misspellings. The rest of them (65%) were not misspellings, but proper nouns, foreign words, derived words, etc. The misspellings were manu­ally revised, and a correction was assigned to each of them based on the context, with the result that the same mis­spelled word could have more than one correction and therefore appear in more than one error pair. Every mis­spelling was classified as belonging to a more general class of errors (i.e. error type). The frequency of each er­ror type and error pair occurrence was calculated.

1 The Original Works Creation (OWC) sites were created to gather unedited text. The participants from Madrid and Mexico City chose from several topics and wrote for about 30 minutes. The configuration of the software and the hardware was set in such a way that authors were only allowed to make minimal im­mediate corrections. The files have only been sentence-separated, and were not spell-checked or grammar-checked.

2 Our goal was to find the most common spelling error patterns in Spanish. For this reason, we started looking for unknown words in the corpus. In this sense, it was not the aim of this study to look for errors in which a correctly spelled word is substituted for another word that exists in the same language (e.g., from-form in English, and mâs-mas in Spanish).

We derived from the corpus an initial classification of 142 error types that exhaustively exploit combinations of the four editing operations. This classification incorpo­rates also multi-error misspellings, position of the error in­side of a word, character distance from the intended word, and edit operation type. Additionally, we have introduced finer-grained categorization on idiosyncratic features like diacritics, space, and case. A new distinction was also made in order to collect frequency information on cogni­tive errors of the phonetic type (substitutions of a phoneti­cally correct but orthographically incorrect letter or se­quence of letters for the intended word) (Kukich, 1992). Table 2 shows some examples from this initial classifica­tion.

verdadramnet — verdadera-mente

This typology shows that a full 51% of the misspell­ings found in our corpus are omissions of a diacritic on a vowel. An analysis of the error rate variation across the different sets of texts ratifies that the high frequency of this error type is not due to a lack of balance among the various texts (Table 3).

Conversely, capitalization at the beginning of the word is remarkably more frequent in one of the sets than in the others (as shown in Table 4).

Regarding cognitive errors, in Spanish this error type mostly consists of substitutions of a phonetically correct but orthographically incorrect sequence of letters (i.e. homophones, mistakes on similar phonetic pairs such as b-v, s-x, c-s, ll-y for instance, as in *archibo-archivo, *estenciôn-extensiôn, *llendo-yendo). This type of error is common in single and multi-error patterns. A breakdown of the total number of cognitive errors found in the corpus shows that unedited sets exhibit more frequently this error type than edited sets (Table 5). Most of cognitive errors are instances of substitutions (* biene-viene), omission (*acer-hacer), and addition (*hacerca-acerca).

Although it could be argued that the cause of the sub­stitution of a b for v could be a keyboard adjacency effect rather than a cognitive error, the overwhelming high fre­quency of this error pair (614; see Table 6 below), in comparison to the lower frequency rate of other adjacent keys (*i-u (51) and *u-i (38), for instance) would remain unexplained.

On the other hand, the source of errors involving dia­critics and case could be considered a misconception on the part of the writer, and, hence, cognitive errors. This fact would imply that a full 63% of the errors found in the corpus would be instances of cognitive errors. This figure suggests that prioritizing some instances of this error type may be very helpful for a spelling correction application for Spanish.

In addition, Figures in Table 2 show that in Spanish omission is the most common spelling error, followed by substitution, addition, and transposition in that order.

The error patterns in Spanish seem to corroborate only partially some of the pattern findings reported by Kukick (1992). Kukick's conclusions are the following:

1. most errors (i.e. roughly 80%) tend to be single in­stances of insertions, deletions, substitutions, or transposi­tions,

2. most errors tend to be within one letter in length of the intended word,

6. phonetic errors are harder to correct because they result in greater distortion of the misspelled string.

In the following sections we revisit these statements and compare them with our data from Spanish.

This paper shows that some of the general statements found in the literature about spelling error patterns apply mainly to English and cannot be extrapolated to other lan­guages. It also reveals the beneficial and productive ef­fects of working with real-life data in the targeted lan­guage for the elaboration of an error typology. This type of data provides clues about which error patterns should be promoted to enhance the generation of suggestions lists for corrections and, if dealing with context-dependent er­rors, to promote those flags related to the most common error patterns. It is obvious that errors involving homoph-ony in Spanish are the most likely contextual errors to be flagged.