The Romanian Corpus Annotated with Verbal Multiword Expressions

Verginica Barbu Mititelu, Mihaela Cristescu, Mihaela OnofreiACL 2019

TL;DR

This paper introduces the first comprehensive Romanian corpus annotated for Verbal Multiword Expressions (VMWEs), following the PARSEME guidelines. This resource offers valuable insights into the frequency, diversity, and structural characteristics of Romanian VMWEs, serving as a crucial foundation for future NLP research and applications in the language.

What Problem Does This Paper Address?

Multiword Expressions (MWEs), particularly Verbal Multiword Expressions (VMWEs), pose a significant challenge for Natural Language Processing (NLP) due to their idiomatic nature, variability, and discontinuities. While international efforts like the PARSEME project have focused on standardizing MWE annotation across languages, a dedicated, linguistically rich corpus for Romanian VMWEs was lacking. This paper addresses this gap by presenting the creation and analysis of such a corpus, enabling a deeper understanding and facilitating the development of automatic MWE identification and classification systems for Romanian.

Methods

The authors followed the PARSEME guidelines to annotate a corpus of Romanian journalistic texts. The annotation process involved sentence splitting, tokenization, part-of-speech tagging, lemmatization, and syntactic annotation using the TTL tool and UDPipe. VMWEs were then identified and classified into categories defined by PARSEME, including Light Verb Constructions (LVCs) and Verbal Idioms (VIDs). A rigorous annotation process, including consistency checks and double-annotation for a subset of the data, was employed to ensure quality.

Datasets

The primary dataset used is a Romanian corpus of 56,703 sentences (1,015,623 tokens) derived from concatenated editions of the Agenda newspaper. This corpus was made publicly available under a CC BY 4.0 license. A subset of 2,503 sentences was double-annotated to assess inter-annotator agreement.

Key Insights and Ideas

  • Romanian VMWEs show unique characteristics: While sharing some similarities with other Romance languages, Romanian VMWEs exhibit distinct patterns in their distribution and types, suggesting a potential influence from Slavic languages.
  • Inherently Reflexive Verbs (IRVs) are highly prevalent: The IRV category, which includes verbs that are rarely used without a reflexive clitic or whose meaning changes significantly with it, is the most represented VMWE type in the Romanian corpus.
  • Verb productivity varies significantly: Certain verbs, like "da" (give) and "pune" (put), are highly productive, appearing in multiple VMWE types, while others are less so. This highlights the role of specific verbs in forming idiomatic expressions.
  • Reduced VMWE length is common: Similar to other languages in the PARSEME initiative, Romanian VMWEs tend to be short, averaging 2.15 words.
  • Structural diversity within VMWE types: While LVCs and VIDs have distinct structural patterns, there's notable variation within these categories, particularly in how nouns and prepositions are integrated with verbs.

Key Results

The annotated Romanian corpus contains 5,891 VMWEs, representing 486 unique expressions. The most frequent VMWE type is IRV (3,784 occurrences), followed by VID (1,611 occurrences). LVC.full and LVC.cause are less frequent, with 313 and 183 occurrences, respectively. The study identified that 122 VMWEs (approximately a quarter) occur only once in the corpus, indicating a significant number of rare expressions. Inter-annotator agreement scores for VMWE span and categorization were moderate (Fspan: 0.533, Kspan: 0.491), suggesting the inherent difficulty of VMWE annotation.

Limitations and Future Work

The authors acknowledge that the findings are specific to the journalistic genre of the corpus and may not represent a general characterization of all Romanian VMWEs. Future work could involve extending the analysis to other text genres, exploring morphological and syntactic properties of head verbs in VMWEs, and leveraging the annotated corpus for comparative linguistic studies across languages. The corpus's availability is expected to foster further research in automatic VMWE identification and classification for Romanian.

Summary updated

BibTeX
@inproceedings{mititelu2019romanian,
  author = {Verginica Barbu Mititelu and Mihaela Cristescu and Mihaela Onofrei},
  title = {{The Romanian Corpus Annotated with Verbal Multiword Expressions}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2019},
}
Figure