Investigating the cross-lingual translatability of VerbNet-style classification.

Area of Science:

• Computational Linguistics
• Lexicography
• Natural Language Processing (NLP)

Background:

• VerbNet is an extensive English verb lexicon valuable for NLP tasks.
• Multilingual NLP is hindered by limited availability of verb classifications in languages other than English.
• Manual development of VerbNet is resource-intensive, prompting research into translation approaches.

Purpose of the Study:

• To systematically investigate the applicability and accuracy of translating VerbNet classes across typologically diverse languages.
• To develop and apply a method for translating VerbNet classes from English to Polish, Croatian, Mandarin, Japanese, Italian, and Finnish.
• To assess the cross-lingual potential of VerbNet classes and provide gold standards for automatic verb classification.

Main Methods:

• Developed a systematic method for translating VerbNet classes from English.
• Applied the translation method to Polish, Croatian, Mandarin, Japanese, Italian, and Finnish.
• Evaluated translatability, inter-annotator agreement, and overlap in classifications.

Main Results:

• Achieved high translatability for Polish (96% of English member verbs translated).
• Demonstrated strong inter-annotator agreement and promising overlap in Polish classifications.
• Obtained equally promising results for Croatian, Mandarin, Japanese, Italian, and Finnish.

Conclusions:

• VerbNet classes exhibit significant cross-lingual potential.
• The proposed translation method is effective for creating multilingual verb classifications.
• The study provides valuable resources (annotation guidelines and six language-specific verb classifications) for NLP research.