Explorations of the Problem of Broad-coverage and General Domain Event Analysis: The Estonian Experience
Date
2016-11-25
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Seoses tekstide suuremahulise digitaliseerimisega ning digitaalse tekstiloome järjest laiema levikuga on tohutul hulgal loomuliku keele tekste muutunud ja muutumas masinloetavaks. Masinloetavus omab potentsiaali muuta tekstimassiivid inimeste jaoks lihtsamini hallatavaks, nt lubada rakendusi nagu automaatne sisukokkuvõtete tegemine ja tekstide põhjal küsimustele vastamine, ent paraku ei ulatu praegused automaatanalüüsi võimalused tekstide sisu tegeliku mõistmiseni. Oletatakse, tekstide sisu mõistvale automaatanalüüsile viib meid lähemale sündmusanalüüs – kuna paljud tekstid on narratiivse ülesehitusega, tõlgendatavad kui „sündmuste kirjeldused”, peaks tekstidest sündmuste eraldamine ja formaalsel kujul esitamine pakkuma alust mitmete „teksti mõistmist” nõudvate keeletehnoloogia rakenduste loomisel.
Käesolevas väitekirjas uuritakse, kuivõrd saab eestikeelsete tekstide sündmusanalüüsi käsitleda kui avatud sündmuste hulka ja üldvaldkonna tekste hõlmavat automaatse lingvistilise analüüsi ülesannet. Probleemile lähenetakse eesti keele automaatanalüüsi kontekstis uudsest, sündmuste ajasemantikale keskenduvast perspektiivist. Töös kohandatakse eesti keelele TimeML märgendusraamistik ja luuakse raamistikule toetuv automaatne ajaväljendite tuvastaja ning ajasemantilise märgendusega (sündmusviidete, ajaväljendite ning ajaseoste märgendusega) tekstikorpus; analüüsitakse korpuse põhjal inimmärgendajate kooskõla sündmusviidete ja ajaseoste määramisel ning lõpuks uuritakse võimalusi ajasemantika-keskse sündmusanalüüsi laiendamiseks geneeriliseks sündmusanalüüsiks sündmust väljendavate keelendite samaviitelisuse lahendamise näitel.
Töö pakub suuniseid tekstide ajasemantika ja sündmusstruktuuri märgenduse edasiarendamiseks tulevikus ning töös loodud keeleressurssid võimaldavad nii konkreetsete lõpp-rakenduste (nt automaatne ajaküsimustele vastamine) katsetamist kui ka automaatsete märgendustööriistade edasiarendamist.
Due to massive scale digitalisation processes and a switch from traditional means of written communication to digital written communication, vast amounts of human language texts are becoming machine-readable. Machine-readability holds a potential for easing human effort on searching and organising large text collections, allowing applications such as automatic text summarisation and question answering. However, current tools for automatic text analysis do not reach for text understanding required for making these applications generic. It is hypothesised that automatic analysis of events in texts leads us closer to the goal, as many texts can be interpreted as stories/narratives that are decomposable into events. This thesis explores event analysis as broad-coverage and general domain automatic language analysis problem in Estonian, and provides an investigation starting from time-oriented event analysis and tending towards generic event analysis. We adapt TimeML framework to Estonian, and create an automatic temporal expression tagger and a news corpus manually annotated for temporal semantics (event mentions, temporal expressions, and temporal relations) for the language; we analyse consistency of human annotation of event mentions and temporal relations, and, finally, provide a preliminary study on event coreference resolution in Estonian news. The current work also makes suggestions on how future research can improve Estonian event and temporal semantic annotation, and the language resources developed in this work will allow future experimentation with end-user applications (such as automatic answering of temporal questions) as well as provide a basis for developing automatic semantic analysis tools.
Due to massive scale digitalisation processes and a switch from traditional means of written communication to digital written communication, vast amounts of human language texts are becoming machine-readable. Machine-readability holds a potential for easing human effort on searching and organising large text collections, allowing applications such as automatic text summarisation and question answering. However, current tools for automatic text analysis do not reach for text understanding required for making these applications generic. It is hypothesised that automatic analysis of events in texts leads us closer to the goal, as many texts can be interpreted as stories/narratives that are decomposable into events. This thesis explores event analysis as broad-coverage and general domain automatic language analysis problem in Estonian, and provides an investigation starting from time-oriented event analysis and tending towards generic event analysis. We adapt TimeML framework to Estonian, and create an automatic temporal expression tagger and a news corpus manually annotated for temporal semantics (event mentions, temporal expressions, and temporal relations) for the language; we analyse consistency of human annotation of event mentions and temporal relations, and, finally, provide a preliminary study on event coreference resolution in Estonian news. The current work also makes suggestions on how future research can improve Estonian event and temporal semantic annotation, and the language resources developed in this work will allow future experimentation with end-user applications (such as automatic answering of temporal questions) as well as provide a basis for developing automatic semantic analysis tools.