Evaluating Bilingual Lexicon Induction without Lexical Data

Kód výsledku v IS VaVaI

<a href="https://www.isvavai.cz/riv?ss=detail&h=RIV%2F00216224%3A14330%2F25%3A00142131" target="_blank" >RIV/00216224:14330/25:00142131 - isvavai.cz</a>

Výsledek na webu

<a href="https://acl-bg.org/proceedings/2025/RANLP%202025/pdf/2025.ranlp-1.34.pdf" target="_blank" >https://acl-bg.org/proceedings/2025/RANLP%202025/pdf/2025.ranlp-1.34.pdf</a>

DOI - Digital Object Identifier

<a href="http://dx.doi.org/10.26615/978-954-452-098-4-034" target="_blank" >10.26615/978-954-452-098-4-034</a>

Jazyk výsledku

angličtina

Název v původním jazyce

Evaluating Bilingual Lexicon Induction without Lexical Data

Popis výsledku v původním jazyce

Bilingual Lexicon Induction (BLI) is a fundamental task in cross-lingual word embedding (CWE) evaluation, aimed at retrieving word translations from monolingual corpora in two languages. Despite the task’s central role, existing evaluation datasets based on lexical data often contain biases such as a lack of morphological diversity, frequency skew, semantic leakage, and overrepresentation of proper names, which undermine the validity of reported performance. In this paper, we propose a novel, language-agnostic evaluation methodology that entirely eliminates the dependency on lexical data. By training two sets of monolingual word embeddings (MWEs) using identical data and algorithms but with different weight initialisations, we enable the assessment on the BLI task without being affected by the quality of the evaluation dataset. We evaluate three baseline CWE models and analyse the impact of key hyperparameters. Our results provide a more reliable and bias-free perspective on CWE models’ performance.

Název v anglickém jazyce

Evaluating Bilingual Lexicon Induction without Lexical Data

Popis výsledku anglicky

Bilingual Lexicon Induction (BLI) is a fundamental task in cross-lingual word embedding (CWE) evaluation, aimed at retrieving word translations from monolingual corpora in two languages. Despite the task’s central role, existing evaluation datasets based on lexical data often contain biases such as a lack of morphological diversity, frequency skew, semantic leakage, and overrepresentation of proper names, which undermine the validity of reported performance. In this paper, we propose a novel, language-agnostic evaluation methodology that entirely eliminates the dependency on lexical data. By training two sets of monolingual word embeddings (MWEs) using identical data and algorithms but with different weight initialisations, we enable the assessment on the BLI task without being affected by the quality of the evaluation dataset. We evaluate three baseline CWE models and analyse the impact of key hyperparameters. Our results provide a more reliable and bias-free perspective on CWE models’ performance.

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

S - Specificky vyzkum na vysokych skolach

Rok uplatnění

2025

Kód důvěrnosti údajů

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

Název statě ve sborníku

Proceedings of the 15th International Conference on Recent Advances in Natural Language Processing (RANLP)

ISBN

9789544520984

ISSN

2603-2813

e-ISSN

—

Počet stran výsledku

8

Strana od-do

275-282

Název nakladatele

Incoma Ltd.

Místo vydání

Varna, Bulgaria

Místo konání akce

Varna

Datum konání akce

1. 1. 2025

Typ akce podle státní příslušnosti

WRD - Celosvětová akce

Kód UT WoS článku

—