GiellaLT provides an infrastructure for rule-based language technology aimed at minority and indigenous languages, and streamlines building anything from keyboards to speech technology.
2012-2013: LA-2012_1a-31112, Feilkorpus for å testa stavekontrollar for grønlandsk, islandsk, lulesamisk og nordsamisk
| Language/role | Project member |
|---|---|
| ISL | Hulda Óladóttir |
| KAL | Elin Neshamar |
| SME | Thomas Omma |
| SMJ | Inga Mikkelsen |
| Project lead & test bench development | Sjur Moshagen |
Maja Kappfjell has worked with SMA, but is formally not part of the project. We have tried to keep the SMA corpus in synch with the rest, though, in terms of markup standards and reaching a final state at the same time as the rest of the languages.
| Language | No Typos | No of running words | % typos of all words |
|---|---|---|---|
| ISL | 915 | 149 532 | 0,61 % |
| KAL | 224 | 21 723 | 1,03 % |
| SMA | 9074 | 41 682 | 21,77 % |
| SME | 2370 | 47 293 | 5,01 % |
| SMJ | 1170 | 22 536 | 5,19 % |
There is a very big span in % of errors in the languages, from ISL 0,61% to SMA 21,77%. It is quite obvious, though, that in the SMA case, the main issue is one of norm:
The SMA norm prescribes the use of Swedish ö and Norwegian æ - a combination that is not easily available on any standard keyboard (unless you really search for it). This is causing most writers to just ignore this aspect of the norm, and they use either both Norwegian æ and ø all the time, or both Swedish ä and ö all the time. When the sounds these letters represent are also quite frequent, it leads to a situation where every sentence contains spelling errors.
For all languages we tried to collect texts that would be representative of spelling errors made by native speakers - because that’s the main target user group for the spellers. At the same time the texts should be relavitely easy to collect - corpus text collection can be a very time consuming activity. What we ended up with was the following main text categories:
Most texts are available in an open-access corpus repository, but some of the texts are stored in a closed repository for copyright and privacy issues.
The testing done so far is only done on the freely available texts.
| ISL | Precision: 9,5% | Recall: 71,73% | Accuracy: 95,7% |
| KAL | Precision: 5,35% | Recall: 82,59% | Accuracy: 84,76% |
| SMA | Precision: 85,64% | Recall: 94,53% | Accuracy: 95,36% |
| SME | Precision: 72,88% | Recall: 91,05% | Accuracy: 97,85% |
| SMJ | Precision: 54,57% | Recall: 89,74% | Accuracy: 95,59% |
| Language | Speed | Memory use |
|---|---|---|
| ISL | 37,31 words/second | 17 172 Kb |
| KAL | 990,2 words/second | 58 952 Kb |
| SMA | 24,11 words/second | 4 480 Kb |
| SME | 76,05 words/second | 6 380 Kb |
| SMJ | 112,03 words/second | 5 104 Kb |