G55G75G72G3G83G72G85G73G82G85G80G68G81G70G72G3G82G73G3G68G3G74G85G68G80G80G68G85G3G70G75G72G70G78G72G85G3G90G76G87G75G3G71G72G89G76G68G81G87G3G79G68G81G74G88G68G74G72G3G76G81G83G88G87G3
 
Janne Bondi Johannessen, Kristin Hagen and Pia Lane  
The Text Laboratory, University of Oslo  
Pb 1102 Blindern 
0317 Oslo, Norway 
{j.b.johannessen, kristin.hagen, p.m.j.lane}@ilf.uio.no 
 
G36G69G86G87G85G68G70G87c3
This paper reports on an evaluation performed 
on the Grammar Checker for Norwegian 
(NGC), developed at The Text Laboratory, 
University of Oslo.
1
 The ability of the NGC to 
find errors made by different “non-standard” 
linguistic groups is analysed and compared to 
its performance when tested on texts written 
by “standard” users. Then possible ways of 
adapting the NGC for use on deviant language 
input are discussed. 
G20G17G3G44G81G87G85G82G71G88G70G87G76G82G81G3
This paper reports on the results of an evaluation 
we have performed on the Grammar Checker for 
Norwegian (NGC), developed at The Text 
Laboratory, University of Oslo. The NGC is 
now part of Microsoft Word in the Office XP 
package released in 2001.
c21
  The goal of the NGC 
was decided partly by that of the Swedish 
Grammar Checker (SGC, Arppe 2000 and Birn 
2000), designed to detect what were assumed to 
be the errors of “standard” users, and partly by a 
wish to include more linguistically advanced 
features. The kind of grammatical mistakes 
made by linguistically “non-standard”
3
 groups 
was not taken into account, and this kind of tool 
obviously would be beneficial to these groups.  
Having provided an overview of the main 
method behind the NGC, we will give a general 
overview of the kinds of errors that the NGC is 
designed to detect. Then we will show how it 
performs on various deviant language input 
                                                   
1
 http://www.hf.uio.no/tekstlab/ 
2
 The NGC was developed  for the Finnish company 
Lingsoft http://www.lingsoft.fi/.   
3
 Non-native spakers, deaf people, aphasics and dyslexics. 
(essays written by Slav and Chinese students, 
and Norwegian deaf children).  
G21G17G3 G55G75G72G3G73G72G68G87G88G85G72G86G3G82G73G3G87G75G72G3G49G42G38G3
G21G17G20G3 G55G75G72G3G80G72G87G75G82G71G3G69G72G75G76G81G71G3G87G75G72G3G49G42G38G3
The NGC was developed using Constraint 
Grammar (Karlsson et al. 1995). Like the SGC 
the NGC has three main parts in addition to an 
initial tokenizer (spell checking is performed at a 
previous stage): 
 
  A morphological analyser (NOBTWOL), 
which provides each word form with all of its 
lexically possible readings (grammatical tags).  
 
  A morphological CG disambiguator, which 
eliminates incorrect tags according to the 
grammatical context (Karlsson et. al 1995, 
Hagen, Johannessen and Nøklestad 2000a and 
2000b). 
 
  An error detector that identifies different kinds 
of grammatical errors.  
 
There is an interesting problem 
regarding the construction of a grammar 
checker. On the one hand it is necessary to have 
as much grammatical information as possible 
about the particular text that is going to be 
checked. On the other hand, it is very difficult to 
perform any such grammatical analysis, since 
grammatical features (“errors”) essential for the 
analysis might be missing. We tried to solve the 
problem by relaxing many of the requirements 
of the disambiguating tagger described above, 
since it was originally developed for 
grammatically correct texts. An example of this 
is the original CG rule assigning a  determiner 
reading to a word that is next to a noun and 
agrees with it in number and gender:                                           
 
(01) (@w =! (det neut) 
 (0 DEF-DET) 
 (*1 DEF-SG-NEUT-NOUN *L) 
 (NOT LR0 NOT-ADJ-NOUN *L) 
 (NOT *L NOT-ADV-ADJ)) 
 
The rule (one of approximately 2000 
rules) says that if a word is definite and has 
neuter determiner as one of its readings, but 
there is a neuter definite singular noun to its 
right, with nothing but adverbs and adjectives in 
between, then the determiner reading is correct. 
This rule ensures that the first word in the 
sentence below is correctly tagged as a 
determiner and not e.g. a pronoun: 
 
c11c19c21c12c3
c39c72c87   eplet       likte han godt 
the.DEF.NEUTER.SG apple.DEF.NEUTER.SG liked he well 
’That apple, he liked well.’ 
 
The tagger can then safely assume that 
whatever does not agree with the noun to its 
right is not part of the same noun phrase, and 
therefore is a pronoun. However, a G74G85G68G80G80G68G85G3
G70G75G72G70G78G72G85 can never assume that anything is 
correct, and cannot rely on the agreement 
features of the determiner and the noun. Instead, 
it ought to be able to detect any missing 
agreement and point out the error. So the new 
relaxed tagger leaves more ambiguity. Instead, 
very specific error rules are introduced in the 
NGC. Rule (03) below (one of 700 error rules) 
detects gender disagreement between a 
determiner and the following noun (04). 
 
 (03) (@w =s! (@ERR) 
(0 DET-DEF-NEUT) 
(NOT -1 DITRANS) 
(1C NOUN-SG-DEF) 
(NOT 1 NEUT) 
(1 MASC)) 
 
(04)  *Jenta så det bilen 
The.girl saw the.DEF.NEUT.SG car.DEF.MASC.SG 
'The girl saw that car.' 
 
This method is reminiscent of that suggested by 
Schneider and McCoy (1998) for their ICICLE 
system designed to help second-language 
learners of English. However, since theirs is a 
grammar based on context-free rules, it is more 
difficult to implement; in order for a parse to be 
successful, all phrases have to be well-formed, 
which means that the grammar must include 
rules for ungrammatical structures. CG has an 
advantage; it does not have to build a full phrase 
structure, thus partial parses are fine, and local 
errors are easily detected. 
 
G21G17G21G3G40G85G85G82G85G3G87G92G83G72G86G3
The NGC detects the following main error types:    
 
 Noun phrase internal agreement:  
Definiteness  c72c87c3c75c88c86c72c87c3c16c16c33c3c72c87c3c75c88c86 
c68c3c3c87c75c72c17c75c82c88c86c72c3c16c16c33c3c17c17c17c3c3c75c82c88c86c72c3
Gender agreement   c72c81c3c75c88c86c3c16c16c33c3c72c87c3c75c88c86 
c68c17c48c36c54c38c17c3c3c75c82c88c86c72c17c49c40c56c55c3c16c16c33c3c68c17c3c49c40c56c55c3c17c17c17c3
Number agreement c72c87c3c72c83c79c72c85c3c16c16c33c3c72c87c3c72c83c79c72 
c3 c3 c68c81c3c68c83c83c79c72c86c3c16c16c33c3c17c17c17c3c68c83c83c79c72c3
 Subject complement agreement  
c37c76c79c72c81c3c72c85c3c85c161c71c87c3c16c16c33c3c37c76c79c72c81c3c72c85c3c85c161c71 
c3c55c75c72c3c70c68c85c17c48c36c54c38c3c76c86c3c85c72c71c17c49c40c56c55c17c16c16c33c3c17c17c17c3c3c85c72c71c17c3c48c36c54c38c3
c135c3Negative polarity itemsc3
c45c72c74c3c78c77c161c83c72c85c3c81c82c72c81c3c69c82c78c16c16c33 c45c72c74c3c78c77c161c83c72c85c3c72c76c3c69c82c78c3c3
c3 c44c3c69c88c92c3c68c81c92c3c69c82c82c78c16c16c33c3c17c17c17c3c3c68c3c69c82c82c78c3
c135c3c82c74c18c110c3errors (conjunction/ inf. marker)c3
c39c72c3c74c76c78c78c3c110c3c86c68c81c74c3c16c16c33c3c39c72c3c74c76c78c78c3c82c74c3c86c68c81c74 
c55c75c72c92c3c90c68c79c78c72c71c3c87c82c3c86c68c81c74c3c16c16c33c3c17c17c17c3c68c81c71c3c86c68c81c74c3
  Too many or no finite verb(s) in a sentence  
c44c3c49c82c85c74c72c3c72c85c3c89c68c85c3c71c72c87c3c86c79c76c78c3c16c16c33 c44c3c49c82c85c74c72c3c72c85c3c71c72c87c3c86c79c76c78c17 
c3 c44c81c3c49c82c85c90c68c92c3c76c86c3c90c68c86c3c76c87c3c86c88c70c75c3c16c16c33c3c17c17c17c3c76c86c3c76c87c3c86c88c70c75 
  Word order errors 
c49c110c3c74c88c87c87c72c81c3c78c82c80c3c16c16c33 c49c110c3c78c82c80c3c74c88c87c87c72c81c17c3
c49c82c90c3c87c75c72c17c69c82c92c3c70c68c80c72c3c16c16c33c3c49c82c90c3c70c68c80c72c3c17c17c17c3
 
G21G17G22G3G3G51G85G72G89G76G82G88G86G3G72G89G68G79G88G68G87G76G82G81G3G82G73G3G87G75G72G3G49G42G38G3G3G3
Our guide line, given to us by Lingsoft, for the 
acceptable number of “false alarms” was 30% 
(70% of all alarms had to report true errors), and 
it performs well within that limit, with a 
precision of 75%
 
(Hagen, Johannessen and Lane 
2001), compared with 70% for the SGC (Birn 
2000). The recall rate for the NGC has not been 
calculated. 
The figures above were calculated on the 
basis of texts written by advanced language 
users - mostly Norwegian and Swedish 
journalists, with few errors in each text. Most of 
the errors were not due to lack of knowledge of 
Norwegian grammar, but rather to modern word 
processing: too quick use of functions like cut 
and paste, insert etc. For example,  two finite 
modal verbs next to each other would not be 
uncommon. However, one would assume that 
less linguistically advanced users might benefit 
more from this kind of tool. In the next sections 
we shall evaluate the NGC on texts produced by 
various non-standard language users. 
 
G22G17G3 G49G72G90G3G72G89G68G79G88G68G87G76G82G81G3G3
G22G17G20G3G40G85G85G82G85G86G3G68G81G71G3G81G82G81G16G86G87G68G81G71G68G85G71G3G79G68G81G74G88G68G74G72G3G88G86G72G85G86G3
We have so far tested four groups of foreign 
students and one group of Norwegian deaf 
pupils, and are in the process of testing aphasics 
and dyslexics. We have divided the errors into 
five groups:      
  G76G12G3 G44G71G76G82G80G68G87G76G70G3 G72G85G85G82G85G86G29 This covers 
language use not strictly speaking 
ungrammatical, just «foreign», G76G76G12G3 G47G72G91G76G70G68G79G3
G72G85G85G82G85G86G29 Wrong word, lack of subcategorised 
word, or a word too many,G3G76G76G76G12G3G54G92G81G87G68G70G87G76G70G3G72G85G85G82G85G86G29G3
Wrong word order, lack of word (that's not 
subcategorised by a particular word), negative 
polarity errors, wrong choice of 
pronoun/anaphor, G76G89G12G3 G48G82G85G83G75G82G79G82G74G76G70G68G79G3 G72G85G85G82G85G86G29G3
Morphological features,  NP agreement 
(number, definiteness, gender), predicative 
agreement, tense of verbs,G3G89G12G3G51G85G68G74G80G68G87G76G70G3G72G85G85G82G85G86G29G3
Errors that involve sentence-external rules: 
Definiteness of NPs (due to known or new 
information), verb tense that ought to follow 
from the context. 
  More specifically, we have tested the 
NGC on essays written by Norwegian deaf 
pupils (11-15 years old) and four groups of 
foreign university students in Norway (Slav and 
Chinese students on Level II (Intermediate) and 
Level III (Advanced). We have included papers 
written by a control group of Norwegian pupils, 
as the student essays were hand written and the 
initial precision of the NGC was calculated on 
word-processed texts. We will also test the NGC 
on essays written by dyslexic and aphasic adults. 
G3
G22G17G21G3G55G72G86G87G3G85G72G86G88G79G87G86G3
There is not enough space to give the individual 
test results here. Let us instead illustrate with 
one group, the Chinese intermediate students. 
There were 15 essays of an average of 300 
words, altogether 4500 words, the same amount 
as for the other test groups. The vast majority of 
the detected errors are morphological ones, see 
table (05): 
 
 
(05) Errors detected by the NGC for Chinese Level II stud. 
c40c85c85c82c85c3c87c92c83c72c86c3 c49c88c80c69c72c85c3c3
Syntactic 4 
Morphological 28 
 
(06) c40c91c68c80c83c79c72c3c82c73c3c86c92c81c87c68c70c87c76c70c3c72c85c85c82c85c29 
 Fordi jeg kan ikke uttrykke meg 
 because I can not express myself 
c214c3Fordi jeg ikke kan uttrykke meg 
 
(07)   c40c91c68c80c83c79c72c3c82c73c3c80c82c85c83c75c82c79c82c74c76c70c68c79c3c72c85c85c82c85c29 
Taiwan er et lite øy 
           Taiwan is a (neut) small (neut)  
island (masc) 
c214c3Taiwan er en liten øy 
 
However, in order to evaluate the NGC 
properly with respect to the Chinese students, 
we have to look at all errors made.  
 
(08) Errors by Chinese Lev. II stud. not found by the NGC : 
c40c85c85c82c85c3c87c92c83c72c86c3 c49c88c80c69c72c85c3c3
Syntactic 68 
Morphological 45 
Lexical 70 
Pragmatic 13 
Idiomatic 32
c54c88c80c29c3 c21c21c27c3
 
In addition to the 32 errors detected by 
the NGC, the Chinese Level II students made 
228 errors that were not detected by the NGC, 
i.e. only 12% were found. But notice that nearly 
half the errors (115) are lexical, idiomatic and 
pragmatic ones – error types that have not even 
been attempted to be detected by the NGC. 
 
(09) c40c91c68c80c83c79c72c3c82c73c3c76c71c76c82c80c68c87c76c70c3c72c85c85c82c85c29 
Nå er jeg i Norge som alle er dyre 
now am I in Norway which all  are expensive (pl) 
c214c3Nå er jeg i Norge hvor alt er dyrt 
 
(10)  c40c91c68c80c83c79c72c3c82c73c3c79c72c91c76c70c68c79c3c72c85c85c82c85c29c3 
Jeg var veldig redd av blod 
        I was very afraid of blood 
c214c3Jeg var veldig redd for blod 
 
(11)  c40c91c68c80c83c79c72c3c82c73c3c83c85c68c74c80c68c87c76c70c3c72c85c85c82c85c29c3 
Det er en vane du må etablere når du var barn 
             It’s a habit you must establish when you were child 
c214c3Det er en vane du må etablere når du er barn 
 
 
Of the morphological mistakes made by 
the  Chinese Level II students, the NGC detected 
28 out of 73, a recall of 38% - considerably 
higher than the results for all categories taken 
together. It can also be improved by adding 
more morphological rules.    
This is similar to the error pattern of all 
the other non-standard language groups we have 
studied so far (Chinese Level III students, two 
levels of Slav students and deaf Norwegian 
pupils). The NGC finds 10% of the total number 
of errors in the essays written by Slav students. 
For the deaf students, the NGC findings rise 
slightly, to 14%. A reason for the higher 
percentage could be that the deaf pupils make 
many morphological mistakes, a feature the 
NGC is designed to detect. For example, these 
pupils typically use non-finite verb forms and 
wrong gender for nouns. 
 
Like the Chinese students, both the 
Slavs and the deaf pupils have a very high 
percentage of «non-grammatical» errors, i.e., 
lexical, idiomatic and pragmatic. The non-
grammatical errors of the Slav students amount 
to 60% of all errors, while the number for the 
deaf pupils is 52%. 
 
However, there are also big differences 
between the groups, see table (12) below. For 
example, the foreign language students have 
fewer idiomatic and pragmatic errors than the 
deaf pupils (20% of all errors versus 31%). This 
aspect is even more striking when we look at the 
pragmatic errors only. The Slav students have 
only 4% pragmatic errors (of all errors). The 
Chinese students have a higher number; 9%. The 
deaf students, however, have 22% pragmatic 
errors.  
 
(12) Errors in % of all errors 
c40c85c85c82c85c3c87c92c83c72c86c3 c38c75c76c81c72c86c72c3 c54c79c68c89c3 c39c72c68c73c3
Syntactic 23 17 15 
Morphological 24 23 37 
Lexical 31 41 17 
Pragmatic 9 4 22 
Idiomatic 12 15 9 
 
The deaf students especially make two 
kinds of pragmatic errors: wrong choice of 
definiteness on the basis of given/new 
information, and wrong use of tense (typically a 
change of tense when none is called for). 
Related to this is the morphological kind of error 
mentioned above: lack of finiteness on verbs. 
These numbers, though interesting, are hardly 
surprising; to some extent they reflect the 
linguistic background of these language users. 
The Norwegian Sign Language and Chinese 
have no morphological verb marking or noun 
marking, while Slavic languages have a complex 
system of verb inflection.  
 
The results for the Norwegian control 
group are predictable. They make no non-
grammatical mistakes, few grammatical 
mistakes
4
, and frequently split compounds 
incorrectly. 16% of their errors were found by 
the NGC – slightly higher than the number for 
the other test groups, but much lower than the 
equivalent number of the SGC wich was 
reported to be 35% (Birn 2000) in Swedish 
newspaper texts. Obviously, the reason for the 
lower number is that the essays by the 
Norwegian pupils are originally written by hand, 
and thus lack easily detectable cut-and-paste and 
our word-processing errors. Our ongoing 
research will show us the results for the other 
"non-standard" language groups. 
 
The NGC gives surprisingly few «false 
alarms» (the precision is 95%, as opposed to 
75% for the newspaper texts) in the texts by  
non-standard language groups, due to the fact 
that their language is very simple, suiting the 
shallow analysis performed by the NGC. The 
precision for the Norwegian control group is 
also high: 87%.   
 
G23G3 G38G68G81G3G87G75G72G3G49G42G38G3G69G72G3G76G80G83G85G82G89G72G71G34G3
With a larger-scale error analysis of authentic 
texts from the non-standard groups a lot of new 
knowledge could be found, which would make a 
good basis for improving the NGC. More 
specifically, since morphological and syntactic 
features are governed by sentence-internal rules, 
a rule-based grammar checker like the NGC 
                                                   
4
 Apart from c82c74c18c110 errors (conjunction and inf.marker– 
notoriously difficult because the pronunciation is the same) 
should be able to account for violations of such 
features. 
 In fact, we have done some testing for 
this purpose, and developed rules for 14 new 
morphological and syntactic error types
5
. One 
new feature is the detection of compound words 
erroneously written as two separate words (G78G85G72G73G87G3
G86G87G88G71G76G72G85 ‘cancer studies’ instead of G78G85G72G73G87G86G87G88G71G76G72G85). 
The results are promising: The overall recall for 
morphological errors for our Chinese students 
now increased from 38% to 53%. 
Detection of non-grammatical errors is  
more difficult to improve. Of course, with an 
advanced lexicon containing e.g. detailed 
subcategorisation information, we might be able 
to find some lexical errors, such as certain 
prepositions after certain verbs. On the other 
hand, a lexicon instead of context, since 
something which looks like a subcategorised 
category may in fact be something else. 
Idiomatic expressions are more of a problem, 
but given a large error corpus, it would be 
possible to extract the most common error types 
and look for these later. Pragmatic errors are 
notoriously difficult. But even here, it would be 
possible to use some system to alarm the user 
whenever, for example, a singular count noun 
occurs without a determiner or two verbs have 
different tense in the same sentence. 
G38G82G81G70G79G88G86G76G82G81G3
The Norwegian Grammar Checker was made for 
native Norwegian speakers. We have tested the 
NGC on texts written by "non-standard" 
language users (Slav and Chinese students of 
Norwegian, and Norwegian native deaf pupils). 
It turned out that a high percentage, as much as 
85-90%, of their mistakes remained undetected.  
However, the picture is not quite as bleak as it 
might seem, since the majority of the errors, 50-
60%, were "non-grammatical", i.e. lexical, 
idiomatic and pragmatic, errors which the NGC 
was not designed to detect.  
 
Does this mean that a grammar checker is of no 
help for these groups? We do not think so; 40-
50% of the errors were grammatical, i.e. subject 
                                                   
5
 This additional testing has been performed in cooperation 
with the project A Grammar Checker for Users with 
Special Needs, headed by Torbjørn Nordgård at NTNU. 
to sentence-internal linguistic rules. Since the 
NGC is rule-based, it should be possible to 
account for many of these kinds of errors. 
Indeed, preliminary tests indicate that this is the 
case. 
 
 G36G70G78G81G82G90G79G72G71G74G72G80G72G81G87G86 
 
Our thanks go to Wenche Vagle (the KAL 
project) for providing Norwegian essays, The 
Department of Norwegian for Foreign Students 
at the UiO, for texts from foreign students, 
Elisabeth Svinndal and Skådalen skole for texts 
from deaf pupils, and to Pål Kristian Eriksen for 
various practical and linguistic help. 
G53G72G73G72G85G72G81G70G72G86G3G3
Arppe, A. 2000. Developing a grammar checker for 
Swedish. In Nordgård, T. (ed.) G49G82G71G68G79G76G71G68G3 G10G28G28G3
G51G85G82G70G72G72G71G76G81G74G86G3 G73G85G82G80G3 G87G75G72G3 G20G21
c87c75
G3G49G82G85G71G76G86G78G72G3
G71G68G87G68G79G76G81G74G89G76G86G87G76G78G78G71G68G74G72G85, Department of Linguistics, 
University of Trondheim, p. 13-27.  
Birn, J. 2000. Detecting grammar errors with 
Lingsoft's Swedish grammar checker. In Nordgård, 
T. (ed.) G49G82G71G68G79G76G71G68G3 G10G28G28G3 G51G85G82G70G72G72G71G76G81G74G86G3 G73G85G82G80G3 G87G75G72G3 G20G21
c87c75
G3
G49G82G85G71G76G86G78G72G3 G71G68G87G68G79G76G81G74G89G76G86G87G76G78G78G71G68G74G72G85, Department of 
Linguistics, University of Trondheim, p. 28-40.  
Hagen, K., J.B. Johannessen and P. Lane. 2001: 
Some problems related to the development of a 
grammar checker. Paper presented at G49G50G39G36G47G44G39G36G3
G10G19G20G15G3G87G75G72G3G21G19G19G20G3G49G82G85G71G76G70G3G38G82G81G73G72G85G72G81G70G72G3G76G81G3G38G82G80G83G88G87G68G87G76G82G81G68G79G3
G47G76G81G74G88G76G86G87G76G70G86G15G3G48G68G92G3G21G20G16G21G21G15G3G21G19G19G20G17 
Hagen, K., J.B. Johannessen and A. Nøklestad. 
2000a. The shortcomings of a tagger.  In Nordgård, 
T (red.) G49G82G71G68G79G76G71G68G3 G10G28G28G3 G51G85G82G70G72G72G71G76G81G74G86G3 G73G82G85G80G3 G87G75G72G3 G20G21G87G75G3
G5G49G82G85G71G76G86G78G72G3 G71G68G87G68G79G76G81G74G89G76G86G87G76G78G78G71G68G74G68G85G5, Department of 
Linguistics, University of Trondheim, p. 66-75. 
Hagen, K., J.B. Johannessen and A. Nøklestad. 
2000b. A Constraint-based Tagger for Norwegian.  
I Lindberg, Carl-Erik and Steffen Nordahl Lund 
(red.): G20G26G87G75G3 G54G70G68G81G71G76G81G68G89G76G68G81G3 G38G82G81G73G72G85G72G81G70G72G3 G82G73G3
G47G76G81G74G88G76G86G87G76G70G86G17G3G50G71G72G81G86G72G3G58G82G85G78G76G81G74G3G51G68G83G72G85G86G3G76G81G3G47G68G81G74G88G68G74G72G3
G68G81G71G3 G38G82G80G80G88G81G76G70G68G87G76G82G81G3 G20G28G15G3 31-48, University of 
Southern Denmark.G3G3 
Karlsson, F., A. Voutilainen, J. Heikkilä and A. 
Anttila. 1995. G38G82G81G86G87G85G68G76G81G87G3G42G85G68G80G80G68G85G17 G36G3G47G68G81G74G88G68G74G72G16
G44G81G71G72G83G72G81G71G72G81G87G3G54G92G86G87G72G80G3G73G82G85G3G51G68G85G86G76G81G74G3G56G81G85G72G86G87G85G76G70G87G72G71G3G55G72G91G87G17 
Mouton de Gruyter, Berlin. 
Schneider, D. and K. McCoy. 1998. Recognizing 
Syntactic Errors in the Writing of Second 
Language Learners. In G38G50G47G44G49G42G16G36G38G47G3 G182G28G27G17G3
G51G85G82G70G72G72G71G76G81G74G86G3G82G73G3G87G75G72G3G38G82G81G73G72G85G72G81G70G72G17 
