Hello everyone! I am currently working on research where the idea is to create a POS tagger for the Serbian language. Each word in a sentence is tagged on a maximum of three levels: word type, word subtype and case.
My data is such that each data has a sentence containing pairs of word/label. E.g:
Osnovi/NN koji/PPN isključuju/VMR postojanje/NA prekršaja/NG Originalna/AGN lista/NN sa/SI podacima/NI o/SL oko/RG 2.000/MC bogatih/AGG Grka/NG koja/PPN je/VAR Atini/NN dostavljena/APN 2010./MO je/VAR nestala/VMP ,/Z prenela/VMP je/VAR tada/RG francuska/AGN novinska/AGN agencija/NN AFP/NN ./Z
Realizcijom/NI ovog/PPG pogona/NG zaposliće/VMF se/PXA 21/MC radnik/NN direktno/RG u/SL proizvodnji/NL i/CC 15/MC radnika/NG kroz/SA razne/AGA vodove/NA kooperacije/NG ./Z
Investiciona/AGN vrednost/NN obog/AGG pogona/NG je/VAR 172.849,00/MC dinara/NG ,/Z a/CC značajna/AGA sredstva/NA izdvojila/VMP je/VAR i/QO Vlada/NN AP/NN Vojvodine/NG ,/Z izjavio/VMP je/VAR nakon/SG otvaranja/NG pogona/NG predsednik/NN izvršnog/AGG Veća/NG AP/NN Vojvodine/NG Dr/NG Bojan/NN Pajtić/NN ./Z
This is a slightly specific classification problem because each word in the sentence is labeled separately, and in order to correctly determine the labels, it is necessary to take the whole context and meaning of the sentence.
Does anyone have any advice on which GPT model to use and how to do fine-tuning with this kind of data?