Bootstrapping the Annotation of UD Learner Treebanks

Kód výsledku v IS VaVaI

<a href="https://www.isvavai.cz/riv?ss=detail&h=RIV%2F00216208%3A11320%2F25%3A8CD8IR7P" target="_blank" >RIV/00216208:11320/25:8CD8IR7P - isvavai.cz</a>

Výsledek na webu

<a href="https://www.scopus.com/inward/record.uri?eid=2-s2.0-85198653384&partnerID=40&md5=cccc8dc5beaefe97387ad08737778c1b" target="_blank" >https://www.scopus.com/inward/record.uri?eid=2-s2.0-85198653384&partnerID=40&md5=cccc8dc5beaefe97387ad08737778c1b</a>

DOI - Digital Object Identifier

—

Jazyk výsledku

angličtina

Název v původním jazyce

Bootstrapping the Annotation of UD Learner Treebanks

Popis výsledku v původním jazyce

Learner data comes in a variety of formats, making corpora difficult to compare with each other. Universal Dependencies (UD) has therefore been proposed as a replacement for the various ad-hoc annotation schemes. Nowadays, the time-consuming task of building a UD treebank often starts with a round of automatic annotation. The performance of the currently available tools trained on standard language, however, tends to decline substantially upon application to learner text. Grammatical errors play a major role, but a significant performance gap has been observed even between standard test sets and normalized learner essays. In this paper, we investigate how to best bootstrap the annotation of UD learner corpora. In particular, we want to establish whether Target Hypotheses (THs), i.e. grammar-corrected learner sentences, are suitable training data for fine-tuning a parser aimed for original (ungrammatical) L2 material. We perform experiments using English and Italian data from two of the already available UD learner corpora. Our results show manually annotated THs to be highly beneficial and suggest that even automatically parsed sentences of this kind might be helpful, if available in sufficiently large amounts. © 2024 ELRA Language Resource Association: CC BY-NC 4.0.

Název v anglickém jazyce

Bootstrapping the Annotation of UD Learner Treebanks

Popis výsledku anglicky

Learner data comes in a variety of formats, making corpora difficult to compare with each other. Universal Dependencies (UD) has therefore been proposed as a replacement for the various ad-hoc annotation schemes. Nowadays, the time-consuming task of building a UD treebank often starts with a round of automatic annotation. The performance of the currently available tools trained on standard language, however, tends to decline substantially upon application to learner text. Grammatical errors play a major role, but a significant performance gap has been observed even between standard test sets and normalized learner essays. In this paper, we investigate how to best bootstrap the annotation of UD learner corpora. In particular, we want to establish whether Target Hypotheses (THs), i.e. grammar-corrected learner sentences, are suitable training data for fine-tuning a parser aimed for original (ungrammatical) L2 material. We perform experiments using English and Italian data from two of the already available UD learner corpora. Our results show manually annotated THs to be highly beneficial and suggest that even automatically parsed sentences of this kind might be helpful, if available in sufficiently large amounts. © 2024 ELRA Language Resource Association: CC BY-NC 4.0.

Druh

D - Stať ve sborníku

CEP obor

—

OECD FORD obor

10201 - Computer sciences, information science, bioinformathics (hardware development to be 2.2, social aspect to be 5.8)

Projekt

—

Návaznosti

—

Rok uplatnění

2024

Kód důvěrnosti údajů

S - Úplné a pravdivé údaje o projektu nepodléhají ochraně podle zvláštních právních předpisů

Název statě ve sborníku

Workshop Build. Using Comp. Corpora, BUCC LREC-COLING - Proc.

ISBN

978-249381431-9

ISSN

—

e-ISSN

—

Počet stran výsledku

7

Strana od-do

111-117

Název nakladatele

European Language Resources Association (ELRA)

Místo vydání

—

Místo konání akce

Torino, Italia

Datum konání akce

1. 1. 2025

Typ akce podle státní příslušnosti

WRD - Celosvětová akce

Kód UT WoS článku

—