From Language to Learning: An LLM-Driven Framework for Multilingual Educational Content

This thesisintroduces "FromLanguagetoLearning," a unified,LLM-driven frameworkdesignedtoaddressthecriticalchallengeofcreatingscalable,personalized educationalcontentacrossdiverselinguisticandpedagogicalcontexts.Itproposesa systematic methodologyforadaptinggeneral-purposelargelanguagemodels(LLMs) into specialized,open-sourceeducationalaids,particularlyforunder-resourcedlan- guages. The cornerstoneofthisframeworkisatask-specificadaptationofdistillationforcre- ating high-quality,task-specificinstructionaldatasets—the Instruct series. First, a state-of-the-artLLMispromptedtogeneratestructurededucationalmaterialsfrom raw,unstructuredtext.Thissyntheticdataisthenusedtofine-tunesmaller,open- sourcemodels,efficientlytransferringsophisticatedcapabilitiesforspecificlearning objectives. The framework’spedagogicalversatilityisdemonstratedthroughaprogressionofap- plications.Tofosterengagementandvocabularyreinforcement,wedevelopedsys- tems forgeneratingeducationalcrosswordpuzzlesin English, Italian,Arabic,and Turkish. Evolvingfromgamificationtoformalassessment,theframeworkwasthen extendedtocreatemodelsforgeneratingmultiple-choiceandshort-answerquestions in Persian andTurkish. Finally,toprovidedirectpedagogicalintervention,the frameworkwasappliedtothenuancedtaskofdeliveringautomatedsyntaxfeedback on studentessaysin English. The efficacyofthisapproachisvalidatedthroughsystematichumanandautomatic evaluations foreachlanguageandtask.Theresultsconsistentlyandsignificantly show thatfine-tuningsmallerLLMsonourgenerated Instruct datasets makes them highlyeffectiveandreliablespecializedtools.Ultimately,thisthesiscontrib- utes ascalableandvalidatedblueprintthatbridgesthegapbetweengeneral-purpose language modelsandthespecificneedsofmultilingualeducation,democratizingthe creationofnext-generation,AI-poweredlearningtechnologies.Acrossthisthesis,we produced 11 publications,released 11 datasets (including 7 in theInstructseries), and trained/evaluated 29 fine-tunedgeneratormodelsacross 5 languages, enabling reproducible,multilingualeducationalNLPatscale.Allcode,datasets,andmod- els developedinthisthesisarepubliclyavailabletotheresearchcommunityat: https: //github.com/KamyarZeinalipour/From-Language-to-Learning