TL;DR
This paper presents a comparative analysis of verbal multiword expressions (VMWEs) in Bulgarian and Romanian wordnets, leveraging the PARSEME corpus and guidelines. It highlights similarities and differences in how these languages lexicalize VMWEs, offering insights for cross-linguistic annotation and resource development.
What Problem Does This Paper Address?
Verbal multiword expressions (VMWEs) are notoriously challenging for natural language processing (NLP) due to their idiomatic nature and varied linguistic structures. While resources like WordNet aim to capture lexical knowledge, the representation and cross-linguistic comparability of VMWEs remain an open problem. This paper addresses this by comparing the VMWEs present in the Bulgarian (BulNet) and Romanian (RoWN) wordnets, which were developed using the standardized PARSEME annotation guidelines. The goal is to understand how these two languages, belonging to different language families (Slavic and Romance), lexicalize similar concepts using VMWEs, and to identify commonalities and divergences that can inform future NLP resource creation and annotation efforts.
Methods
The authors leverage the PARSEME corpus, a multilingual collection of annotated VMWEs, and the Bulgarian and Romanian wordnets (BulNet and RoWN) as their primary data sources. The PARSEME annotation guidelines, which define universal categories for VMWEs like Light Verb Constructions (LVCs) and Verbal Idioms (VIDs), provide a common framework for analysis. The study involves:
- Cross-linguistic comparison: Analyzing the distribution and types of VMWEs in both wordnets and comparing them to each other and to the PARSEME corpus.
- Lexicalization analysis: Investigating how concepts are expressed as VMWEs or single words in each language.
- Identification of correspondences and mismatches: Pinpointing synsets that contain VMWEs in one language but not the other, or where the VMWE type differs.
- Analysis of non-VMWE multitoken phrases: Examining how phrases that don't fit VMWE categories are handled, often to compensate for lexical gaps.
Datasets
- Bulgarian WordNet (BulNet): A lexical database for Bulgarian.
- Romanian WordNet (RoWN): A lexical database for Romanian.
- PARSEME Corpus: A multilingual corpus annotated for verbal multiword expressions, used to inform the wordnet development and analysis. The paper specifically refers to version 1.1 of the PARSEME corpus, which includes subcorpora for 20 languages, with Bulgarian and Romanian being the focus.
Key Insights and Ideas
- Strong Overlap in Lexicalization: Despite language family differences, there's a substantial overlap (40-45%) in synsets containing VMWEs between BulNet and RoWN, suggesting common tendencies in how these languages lexicalize verbal concepts.
- Dominance of Inherently Reflexive Verbs (IRVs): IRVs are the most frequent VMWE type in both wordnets, which is attributed to the semantic nature of reflexive verbs in these languages.
- Bulgarian's Richer LVCs: Bulgarian exhibits a significantly higher number of Light Verb Constructions (LVCs) compared to Romanian, both in the corpus and wordnet. This is partly explained by the broader range of verbs considered "light" in Bulgarian and the language's aspectual system.
- Aspectual Pairs Influence Counts: Bulgarian's aspectual verb pairs (perfective/imperfective) contribute to a higher raw count of VMWEs, as these are sometimes treated as distinct entries. The paper also presents data where these pairs are grouped.
- Lexical Gaps and Compensatory Strategies: Both languages use non-VMWE multitoken phrases (marked as NONE or NO_LEX) to compensate for lexical gaps, often mirroring English single-word or compound lexicalizations.
- Subjectivity in LVC Identification: The identification of LVCs, particularly determining if a verb is "semantically bleached," can be subjective, leading to variations in annotation and potentially affecting cross-linguistic comparisons.
Key Results
- Corpus vs. Wordnet: The distribution of VMWE types in the wordnets generally correlates with their distribution in the respective corpora, confirming that the lexicon reflects language use.
- Quantitative Comparison:
- BulNet contains 92,910 manually verified synsets with 164,418 literals, of which 63,930 are multiword expressions (28.3% of literals).
- RoWN contains 59,348 synsets with 85,277 literals, of which 20,031 are multiword literals (23.5% of literals).
- The intersection of synsets containing VMWEs between BulNet and RoWN comprises 944 synsets.
- VMWE Type Distribution (Table 2):
- IRVs are the most frequent in both wordnets (~56%).
- VIDs are the second most frequent, with a higher percentage in Romanian wordnets (35.1%) compared to Bulgarian (24.0%).
- LVCs are less represented, with a higher percentage in Bulgarian wordnets (13.8% for LVC.full) compared to Romanian (5.8%).
Limitations and Future Work
- Incomplete Wordnets: The authors acknowledge that both BulNet and RoWN are not exhaustive and do not offer a complete picture of the lexical richness of the languages.
- Annotation Conventions: Differences in annotation conventions over time and between teams, particularly for LVCs, can affect comparability.
- Subjectivity in LVC Identification: The judgment of a verb's semantic bleaching for LVC identification is subjective and can lead to variations.
- Future Work: The authors suggest that this comparative analysis can serve as a starting point for drawing conclusions about VMWE scope and distribution, establishing best practices for VMWE description in wordnets, and facilitating future applications like morphosyntactic and inflectional description, and derivational information. Addressing VMWEs in a unified way across wordnets is seen as crucial for improving machine translation and enabling cross-linguistic semantic analysis.
Summary updated