previous next Title Contents

4 Results of testing


We report below the results obtained for the various tests run on the two different products. We have broken down the different tests into sub-fields, according to whether each test deals with common words, proper names, adjectives deriving from a city, numbers in full letters, etc. A separate section will report on lists of words that contain errors.

4.1 Lexical coverage

The lists of words at our disposal were devised from corpora, with no other filtering; hence, they contained proper names, foreign words, etc. We selected two lists to be given to the spellers: one containing only lower-case words (in order to exclude all proper names), derived from the ANSA list, and another larger one derived from the Pisa list, that had not undergone any "cleansing" except the selection of words that had a frequency equal or higher to 20 (the maximum frequency of that list was 545.428).

The reason for choosing this way of proceeding is in the amount of "noise" that each list would inevitably bring, in order to give a likely real-word simulation of a user's spelling checker session. In the ANSA-derived list there would be no proper names, but there would be neologisms, and possibly foreign words. In the construction of the Pisa list, capitalization of words has been ignored, so that no difference is made between strings with upper case and those with lower case letters. Thus for example the strings " con ", " Con ", and " CON " [with] are all counted as instances of the same word and are listed under just one of the possible forms. The choice which form is listed appears to be arbitrary. It was therefore not possible to automatically sort capitalised words as proper names, but we chose to eliminate words with a very low frequency since in that group there would be mainly foreign words, proper names, extremely archaic forms and misspellings.

Apart from the general vocabulary, we submitted to the two checkers lists of proper names: one of personal first names, and another of Italian cities with an administrative function (" capoluogo di provincia "). In order to test the proper " localization " of the spelling checkers, we manually derived the adjectives used for each of the cities, in all the inflected forms. A third list was manually sorted from the Pisa list, consisting of ordinal numbers or age adjectives written in full letters, since some of these have a spelling that might appear unusual for Italian (double vowels). Although ordinal numbers and age numbers in full letters do not have a high frequency in the lists at our disposal, they must be correctly recognized by the checkers.

Throughout this report, the percentages given have been rounded up to the closest integer.

A final word of warning: the two checkers tested seem to treat differently words that are linked with a hyphen, (e.g. " decreto-legge " [legislative decree]) or end with an apostrophe (e.g., " dell' " [of the]) as either one word or two words, so the total word count in the tables below reflects the word count of the checker, and the same list yields a different figure according to the spelling checker used. However, such words occur only in the Pisa derived list, and likewise the different word counts only concern that list.

Lexical coverage

Product: System A





List name
Total no. of words
No. Of words recognised
% of words recognised
Pisa-derived
33377
29961
90
ANSA derived
16527
15518
94

Lexical coverage

Product: System B





List name
Total no. of words
No. Of words recognised
% of words recognised
Pisa-derived
33377
29326
88
ANSA derived
16527
15561
94

Proper names coverage: Personal names

Product: System A



Total no. of words
No. of words recognised
% of words recognised
609
565
93

Proper names coverage: Personal names

Product: System B



Total no. of words
No. of words recognised
% of words recognised
609
484
79

Proper names coverage: Names of cities

Product: System A



Total no. of words
No. of words recognised
% of words recognised
103
103
100

Proper names coverage: Names of cities

Product: System B



Total no. of words
No. of words recognised
% of words recognised
103
102
99

City adjectives

Product: System A



Total no. of words
No. of words recognised
% of words recognised
320
108
34

City adjectives

Product: System B



Total no. of words
No. of words recognised
% of words recognised
320
303
95

Ordinal and age adjectives

Product: System A



Total no. of words
No. of words recognised
% of words recognised
63
18
29

Ordinal and age adjectives

Product: System B



Total no. of words
No. of words recognised
% of words recognised
63
61
97

Lexical coverage




Product

Lists (Pisa and ANSA derived)

% recognized

Proper Names (People and Cities)

% recognized

Adjectives (City, Age and Ordinal)

% recognized

System A
91%
94%
33%
System B
90%
83%
95%

4.2 Error coverage

We report here the results obtained with the ASCC package regarding the behaviour of the two products tested. The errors have been generated using the errgen facility as explained earlier, and the list obtained was submitted to the spelling checkers with the corresponding correct list it was derived from.

We have opted for a mixture of errors that are due to the mechanical mis-typing of a word and the real lack of knowledge of the correct spelling.

Error coverage

Product: System A




Error types
No. of errors

generated

No. of errors

signalled

% of errors

signalled

1. undouble a consonant
4670
4546
97
2. bile-> bbile
233
231
99
3. g insertion in li+vowel
83
82
100
4. cu ->qu
10
9
99
5. qu ->cu
116
115
99
6. cqu ->qu
36
33
92
7. m <->n exchange
105
100
95

Error coverage

Product: System B




Error types
No. of errors

generated

No. of errors

signalled

% of errors

signalled

1. undouble a consonant
4670
4567
98
2. bile-> bbile
233
232
100
3. g insertion in li+vowel
83
83
100
4. cu -> qu
10
10
100
5. qu ->cu
116
116
100
6. cqu ->qu
36
34
94
7. m <->n exchange
105
101
96

Error coverage






Product

Undouble consonant

%

bile ->bbile

%

g + li

%

c/q related errors

%

m/n exchange

%

System A

97

99
100
97
95
System B

98

100
100
99
96

4.3 Suggestion adequacy

We report below the figures regarding the suggestion adequacy of the two products. We remind the reader that for product B the maximum level of suggestion done is 2, due to the unexpected behaviour that that spelling checker showed during the tests. Therefore, in the comparative table, the two products are compared only up to suggestion level 1.

Suggestion adequacy

Product: System A






Error types


No. of errors

recognised

1st sugg. correct

Correct sugg. Among 2nd-5th


Correct sugg. not among first 5


No sugg. offered




No.

%
No.
%
No.
%
No.
%
1. undouble
4546
3028
67
964
21
282
6
272
6
2. bile -> bbile
231
204
88
1
0
3
1
23
10
3. g insertion+

li +Vowel

82
61
74
8
10
5
6
8
10
4. cu -> qu
9
7
78
1
11
0
0
1
11
5. qu -> cu
115
98
85
0
0
3
3
14
12
6. cqu -> qu
33
22
67
9
27
1
3
1
3
7. m<->n
100
48
48
31
31
20
20
1
1

Suggestion adequacy

Product: System B






Error types


No. of errors

recognised

1st sugg. correct

2nd sugg. Correct


Correct sugg. not among first 2


No sugg. offered




No.

%
No.
%
No.
%
No.
%
1. undouble
4567
4055
89
69
2
305
7
138
3
2. bile -> bbile
232
206
89
1
0
12
5
13
6
3. g insertion+

li +Vowel

83
68
82
1
1
8
10
6
7
4. cu -> qu
10
8
80
1
10
1
10
--
--
5. qu -> cu
116
69
59
1
1
5
4
41
35
6. cqu -> qu
34
30
88
1
3
3
9
--
--
7. m<->n
101
48
48
3
3
22
22
28
28

Suggestion adequacy






Product

No. of errors

recognised

1st sugg. correct
2nd sugg. correct
Correct sugg. among

2nd-5th

Sugg. not among first 2/5
No sugg. offered
A

5116

68%
N.A.
20%
(5) 6%

6%

B

5143

87%
2%
N.A.
(2) 7%
4%

previous next Title Contents