You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Known dictionary words are compared with their codeword. Previously, text strings were compared.
Adjusted global StateMap prediction.
ContextMap (HT 128) reduced predictions from 6/4 to 4/3 per context. Use single internal StateMap. All context states are update with that.
ContextMap (HT 32) reduced predictions from 5/4 to 3/2 per context. Removed StateMap based predictions.
StationaryMap for 2 context
In WordsContext use also codeword for dictionary word.
Added SentenceContext for sentance managment. Max 64 sentances (WordsContexts). Search for similarity is performed by compareing codewords (default 53% means match found).
In stemmer add Pronoun word type.
Add InDirectStateMap with order-w mixing of primary predictions (similar to Paq9a/zpaq)
Partial sentance contexts.
Group of SentenceContexts for: lists ('*'), table, wikilinks and regular sentances. Total 4.
Removed SparseMatchModel.
Removed 4 SmallStationaryContextMap contexts
Mixer count from 12 to 24
Added 7 new ContextMap's
Added 22 new InDirectStateMap contexts
Adjusted mixer parameters and contexts
Adjusted ContextMap memory usage
There are 3 mixers layers (+1 in every InDirectStateMap).
For layer 0 mixers about ~40% of updates are skipped.
Some predictions are skipped if line is Category link, after topic 'See also', 'References', 'Bibliography' or 'External links'.
Some low memory ContextMaps are reset after every page (wikipedia article). The StateMap is preserved if it exists.