i find knowledge graphs very interseting : as today the models are making very detailed knowledge graphs which actually are unusable !
so we should be refineing these graphs to contain the specfic entitys we are concerned wiith:
So if we create a knowledge graph of evidence for instance , we may only wish to work with the subtree of Murder Weapo ! .... Or a subtree of a specific person !
this gives i a knowledge about that entity !
SO in truth we do not need the whole tree which can also produce unwanted predictions ( hallecenation ) or in correct prediction due to the data being oversaturated !.
The knowledge graph situation is very valuable to interpret large documents as well as develop large documents and summary ! as given a overal pwerspective of content and the specific entity required a summary can be written which contain the truth inside the tree ... based on the content (rag summarys ) provided !
For me i dont load rag sumarys in to my RAG ! I upload CONTENT! ....
NOw i feel the content actually has no value ! as your only feeding a prompt with some accompaniyingn ststements , which may or not be relevant !
Instead i have devised that saving masked data is better !
SO if we save masked data sentences , in which the entitys are removed ! then we can now save entity trees also ... the model should be able to recontruct information in from the tree with any collection of MAsked sentences !
in fact it is my belief after one conversation with the model that this is actually how data is stored ! << the masked sentences only contain gramatical usage of language , hence any entiityt can be plugged intot the sentences : hence it can produce any past shape with any entity implants !
the boy jumped over the _ or even ___ JUMPED OVER ____ hence any word ca be pla ed in the object or subject location ! it kinda like the snowball algo !
So if we prepend that we should only be saving Masked sentences and knowledge tres about specific entitys then we can always contrcut any sentence ... based on its activty !
because the knowledge tree is what is actuall holding the data ....so based on a masked sentence and a graph or entity tree , we can fid the most probable matches .... SO in this it can produce Highly likely Hallcenation ! as in changing some entitys it becomes truth !
as data and sequeces are past shapes ! it only selects if as the most probable statitical combo !
SO : Refined kowledge graphs containing just the graph associated with the target entity and the ( masked prases which you wish to know information about ) or the likelyhood of these entity tree Fitting these masked sentences ,,
____ KILLED _____ with _____ in the ____
SO given a tre about John in the kithen as the cheif we can see if indeed he fits that patern in the overall knowldge base of entitys and evidence (non masked sentences ) hence we can determine truth by creating hallucenations and checking the statistical value of the constructed sentence !
Hmmmmm Something to think about : in how to actually use these knowledge graphs as in scenarios in which they have actually been used sucessfully the trees were NOT super detailed !
